User portrait generation method and system applied to intelligent health care

Abstract

The invention provides a user portrait generation method and system applied to intelligent health care, and the method comprises the steps: extracting first health care product characterization data of an associated health care product of a target behavior event, and integrating the first health care product characterization data in a same health care service behavior sequence through an identification weight, according to the embodiment of the invention, the second health care product characterization data for more accurately characterizing the target behavior event in the health care service behavior sequence can be obtained, so that the target health care product characterization data is obtained based on the second health care product characterization data fused with the plurality of health care service behavior sequences; taking the health care service behavior characterization data and the target health care product characterization data as a label diffusion target to perform back diffusion on the original portrait label to obtain a target portrait label, generating a target user portrait corresponding to the target user according to the target portrait label, constructing a corresponding user portrait in a behavior sequence dimension of a finer granularity level, and obtaining a user portrait corresponding to the target user. And the accuracy is higher, and the push conversion rate of the health care service based on the user portrait is facilitated.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for generating user portraits for smart health care. Background Art

[0002] With the increasing aging of the population and people's increasing attention to healthy living, the field of health care services has ushered in development opportunities. In order to provide more personalized and accurate health care services, it is particularly important to conduct in-depth analysis and understanding of users' behavioral characteristics, health status and health care needs. In the current field of health care services, user portraits, as an important technical means, are widely used in the prediction of user needs, the allocation of service resources, and the push of personalized services. User portraits collect and analyze various information of users, such as basic information, behavioral habits, health status, etc., to generate a multi-dimensional user feature representation, providing strong support for the decision-making and optimization of health care services. However, traditional user portrait generation methods often have some problems. First, these methods usually construct user portraits based on users' basic information or simple behavioral statistics, and lack in-depth analysis and understanding of users' behavioral sequence data in the process of health care services. Health care service behavior sequence data records a series of behaviors of users in the process of health care services, which is of great significance for in-depth understanding of users' behavior patterns and health care needs. The accuracy of traditional user portrait generation methods needs to be improved. Summary of the invention

[0003] In view of this, the present application provides a method and system for generating user portraits for smart health care. The technical solution of the present application is implemented as follows: On the one hand, the present application provides a user portrait generation method for smart health care, including: obtaining a set of health care service behavior sequences collected for a target user, and extracting health care service behavior representation data of the health care service behavior sequence set; performing behavior detection on the health care service behavior sequences in the health care service behavior sequence set, and obtaining identification weights of behavior identifiers in the health care service behavior sequences, wherein the behavior identifier is used to identify a target behavior event in the health care service behavior sequence, and the identification weight is a parameter for evaluating the predicted probability of the behavior identifier, and is used to indicate the probability of the target behavior event identified by the behavior identifier; extracting relevant information of the target behavior event. The first health care product characterization data of the health care product is integrated, and the first health care product characterization data in the same health care service behavior sequence is integrated by the identification weight to obtain the second health care product characterization data of the health care service behavior sequence, and the second health care product characterization data of multiple health care service behavior sequences are integrated to obtain the target health care product characterization data; any generated original portrait label is obtained, and the original portrait label is reversely diffused with the health care service behavior characterization data and the target health care product characterization data as label diffusion targets to obtain the target portrait label, and the target user portrait corresponding to the target user is generated according to the target portrait label.

[0004] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0005] The present application performs behavior detection on the health care service behavior sequence in the health care service behavior sequence set to obtain the identification weight of the behavior identifier. Because the identification weight can indicate the predicted probability of the behavior identifier, the first health care product characterization data of the health care product associated with the target behavior event is extracted, and the first health care product characterization data in the same health care service behavior sequence is integrated by the identification weight to obtain the second health care product characterization data that more accurately characterizes the target behavior event in the health care service behavior sequence. Based on this, the second health care product characterization data of multiple health care service behavior sequences is integrated to obtain the target health care product characterization data, so that the target health care product characterization data can have sequence-level behavior characteristics, obtain arbitrarily generated original portrait labels, use the health care service behavior characterization data and the target health care product characterization data as label diffusion targets to reversely diffuse the original portrait labels to obtain the target portrait labels, generate the target user portrait corresponding to the target user according to the target portrait label, and construct the corresponding user portrait in the behavior sequence dimension at a finer granularity level, which has higher accuracy and is beneficial to the push conversion rate of health care services based on user portraits.

[0006] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0008] Figure 1 A schematic diagram of the implementation flow of a method for generating a user portrait for smart health care provided in an embodiment of the present application.

[0009] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing this application and are not intended to limit this application.

[0012] The embodiment of the present application provides a method for generating a user portrait for smart health care, which can be executed by a processor of a computer system. The computer system may refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0013] Figure 1 A schematic diagram of the implementation flow of a method for generating a user portrait for smart health care provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes: Step S100: Acquire a set of health care service behavior sequences collected for a target user, and extract health care service behavior representation data of the health care service behavior sequence set.

[0014] The computer system collects the target user's behavior data in health care services from various data sources. These data sources may include, but are not limited to, the use records of the health care service platform, the monitoring data of wearable devices, the health records of medical institutions, and the health questionnaires actively filled out by users. For example, in a smart health care community, the computer system can obtain the records of the target user's use of various health care services in the past period of time from the community's health management system, such as regular physical examinations, participation in rehabilitation training courses, and use of health monitoring equipment (such as smart bracelets, smart blood pressure monitors, etc.). These records are arranged in chronological order to form a set of health care service behavior sequences. Assume that the health care service behavior sequence of the target user A in a week may be: 30 minutes of aerobic exercise on Monday, blood pressure measurement on Tuesday, attending a mental health lecture on Wednesday, 20 minutes of yoga on Thursday, and monitoring sleep quality with a smart bracelet on Friday.

[0015] The purpose of extracting health care service behavior representation data is to convert the original health care service behavior data into a form that can be understood and processed by the computer system for subsequent analysis and calculation. This process usually involves data preprocessing, feature extraction, and feature representation.

[0016] The computer system first cleans the collected raw data to remove noise data, missing values ​​and outliers. For example, when processing the user's blood pressure measurement data, if a measurement value is found to be significantly deviated from the normal range (such as systolic blood pressure exceeding 200mmHg), the computer system can regard it as an outlier and correct or delete it.

[0017] In order to eliminate the dimensional differences between different data sources and different types of data, the computer system standardizes the data. Common standardization methods include Z-score standardization and Min-Max standardization. For example, for the user's exercise time data (unit: minutes) and sleep quality score data (range: 0-10), Z-score standardization can be used to convert them into standard normal distribution data with a mean of 0 and a standard deviation of 1. The computer system can calculate various statistical characteristics of the health care service behavior sequence, such as the mean, median, standard deviation, maximum value, minimum value, etc. For example, for the user's exercise time data, the computer system can calculate the mean, standard deviation and other statistical characteristics of the exercise time every day in a week, which can reflect the user's exercise habits and regularities.

[0018] For some discrete health care service behavior data, the computer system can calculate its frequency of occurrence. For example, in the records of rehabilitation training courses attended by users, the computer system can count the frequency of attendance at each course (such as physical therapy, rehabilitation gymnastics, etc.), and these frequency characteristics can reflect the user's preference for different rehabilitation training courses.

[0019] Since the health care service behavior data is arranged in chronological order, the computer system can extract some time series features to reflect the changing trend and periodicity of the behavior. For example, for the user's sleep quality data, the computer system can calculate the difference in sleep quality scores between two consecutive days to reflect the changing trend of sleep quality; it can also extract the periodic characteristics of sleep quality data through methods such as Fourier transform, such as whether there are fixed days a week with poor sleep quality.

[0020] The computer system represents the extracted health care service behavior representation data in vector form for subsequent calculation and analysis. For example, for a health care service behavior representation data containing multiple features (such as the average exercise duration, the average blood pressure, the average sleep quality score, etc.), it can be represented as an n-dimensional vector.

[0021] For some health care service behavior data with semantic information (such as rehabilitation training course names, health lecture topics, etc.), computer systems can use embedding technology to convert them into low-dimensional vector representations. For example, word vector models (such as Word2Vec, GloVe, etc.) can be used to convert rehabilitation training course names into fixed-dimensional vectors, which can capture the semantic similarities between different course names.

[0022] Step S200: Perform behavior detection on the health care service behavior sequence in the health care service behavior sequence set, and obtain the identification weight of the behavior identifier in the health care service behavior sequence, wherein the behavior identifier is used to identify the target behavior event in the health care service behavior sequence, and the identification weight is a parameter for evaluating the predicted probability of the behavior identifier, which is used to indicate the probability of the target behavior event identified by the behavior identifier.

[0023] When the computer system performs this step, it conducts a detailed analysis of each health care service behavior sequence to determine the identification weight of the behavior identifier. The behavior identifier is a key element specifically used to identify the target behavior event in the health care service behavior sequence, and its identification weight is an important parameter for evaluating the prediction probability of the identifier, which intuitively reflects the probability of the occurrence of the target behavior event pointed to by the behavior identifier.

[0024] In the process of behavior detection, the computer system can use the classification algorithm in machine learning to identify the target behavior event. Taking the support vector machine (SVM) algorithm as an example, the algorithm separates data points of different categories by constructing a hyperplane, thereby realizing the classification of data. For each behavior data point in the health care service behavior sequence, the computer system extracts its features, forms a feature vector, and inputs it into the trained SVM model for classification judgment. Assuming that the target behavior event of "performing high-intensity exercise" is to be identified, the computer system will extract the features related to exercise intensity in the behavior sequence, such as heart rate during exercise, exercise duration, calories consumed by exercise, etc., to form a feature vector x=(x1,x2,x3), where x1 represents heart rate, x2 represents exercise duration, and x3 represents calories consumed. Then the feature vector is input into the SVM model, and the model judges based on the learned classification boundary and outputs the classification result of whether the behavior belongs to "performing high-intensity exercise".

[0025] As for the determination of the identification weight, the computer system can use the probability model to calculate. For example, the naive Bayes algorithm is used, which calculates the probability based on the Bayes theorem and the feature conditional independence assumption. Assume that there are multiple features F1, F2, ..., F n , and the target behavior event T, the naive Bayes algorithm calculates the probability P(T|F1,F2,…,F n ) to determine the identification weight. According to Bayes' theorem, there is a formula: ; Among them, P(F1,F2,…,F n |T) is the joint probability of each feature appearing under the condition that the target behavior event T occurs; P(T) is the prior probability of the target behavior event T occurring; P(F1, F2, …, F n ) is the joint probability of each feature appearing. The computer system calculates these probability values ​​by learning and counting a large amount of historical data, thereby obtaining the identification weight.

[0026] Assume that the target behavior event of interest is "users frequently use a certain rehabilitation device within a specific time period". The computer system first collects the behavioral sequence data of users using rehabilitation devices, including information such as the time, frequency, and duration of device use. Then, the system extracts the relevant features of these behavioral sequences, such as the number of times the device is used per day, the distribution of usage time, etc., to form a feature vector. Next, the above-mentioned probability model is used to calculate the probability of the target behavior event "users frequently use a certain rehabilitation device within a specific time period" occurring under these feature conditions. This probability value is the identification weight of the behavior identifier.

[0027] In actual operation, the computer system can also use deep learning models to improve the accuracy of behavior detection and the precision of identification weight calculation. For example, use recurrent neural networks (RNN) and their variants, such as long short-term memory networks (LSTM) or gated recurrent units (GRU). These models can handle sequence data well and capture the temporal dependencies in behavior sequences. Taking LSTM as an example, it can effectively handle long-term dependency problems by introducing memory units and gating mechanisms. When processing health care service behavior sequences, LSTM can process the behavior data at each time step, remember important information, and make decisions based on the current input and memory status. The computer system inputs the feature vector of the health care service behavior sequence into the LSTM model according to the time step. After a series of calculations and learning, the model outputs the classification results of the target behavior event and the corresponding identification weight.

[0028] Through the use of the above-mentioned multiple implementation methods, the computer system can conduct in-depth analysis and detection of each behavior sequence in the set of health care service behavior sequences, accurately obtain the identification weight of the behavior identifier, and provide strong support for the subsequent extraction of related health care product representation data of the target behavior event and the further generation of user portraits. This makes the generated user portraits more accurate in the behavior sequence dimension at a finer granularity level, which helps to improve the conversion rate of health care service push based on user portraits.

[0029] Step S300: extract the first health care product representation data of the health care product associated with the target behavior event, integrate the first health care product representation data in the same health care service behavior sequence by identifying the weight, obtain the second health care product representation data of the health care service behavior sequence, and integrate the second health care product representation data of multiple health care service behavior sequences to obtain the target health care product representation data.

[0030] First, the computer system extracts the first health care product representation data of the health care product associated with the target behavior event. The target behavior event is a behavior with specific meaning and value in the health care service behavior sequence. For example, in the health care service behavior sequence, "doing aerobic exercise more than three times a week" can be regarded as a target behavior event. The associated health care products are various health care products related to the target behavior event, such as sportswear, sports shoes, smart sports bracelets, etc. for aerobic exercise. In order to extract the first health care product representation data, the computer system needs to analyze and extract the various attributes and features of these associated health care products. For example, for the health care product of smart sports bracelet, its features may include brand, function (such as heart rate monitoring, motion trajectory recording, etc.), price, etc. The computer system can extract these features through a variety of implementation methods, such as natural language processing technology. If the product information exists in text form, the system can convert the product description text into a vector representation through a word vector model, which is used as part of the first health care product representation data; for numerical features, such as price, it can be directly standardized and used as part of the representation data. Assume that the brand of the smart sports bracelet can be represented as a d-dimensional vector through a word vector. , the function is represented by encoding as an f-dimensional vector , the price is expressed as a value p after normalization, then the first health care product characterization data of the smart sports bracelet can be expressed as a vector .

[0031] Next, the computer system will integrate the first health care product characterization data in the same health care service behavior sequence through the identification weight to obtain the second health care product characterization data of the health care service behavior sequence. The identification weight is obtained in step S200 and is used to indicate the probability of the target behavior event identified by the behavior identifier. For example, in the above-mentioned target behavior event of "performing aerobic exercise more than three times a week", it is assumed that the identification weight of its behavior identifier is w. In the same health care service behavior sequence, there may be multiple first health care product characterization data of related health care products, such as characterization data of sportswear and sports shoes in addition to smart sports bracelets. Etc. The computer system can be integrated by weighted summation, and the calculation formula is as follows: ;in, represents the second health care product characterization data, w i represents the identification weight corresponding to the i-th associated health care product (in the same behavior sequence, the weights for different products may be different), The first health care product representation data of the i-th related health care product. For example, if the first health care product representation data of the smart sports bracelet , identification weight w1=0.3; the first health care product characterization data of sportswear , identification weight w2=0.2; the first health care product characterization data of sports shoes , the identification weight w3=0.5, then the second health care product characterization data .

[0032] Finally, the computer system integrates the second health care product representation data of multiple health care service behavior sequences to obtain the target health care product representation data. In actual smart health care scenarios, a target user may have multiple health care service behavior sequences, such as behavior sequences in different time periods. Assuming that there are m health care service behavior sequences, the corresponding second health care product representation data are The computer system can be integrated by using the average method or the weighted average method. Taking the average method as an example, the calculation formula is: ; Among them, \vec{x_{target}} represents the target health care product representation data. In this way, the computer system can integrate the information of multiple health care service behavior sequences to obtain more comprehensive and accurate target health care product representation data. For example, in one month, the target user has four different weeks of health care service behavior sequences, and the corresponding second health care product representation data is obtained respectively. , then the target health care product characterization data .

[0033] Through the above series of operations, the computer system can extract representation data from the health care products associated with the target behavior events, and through identification weights and integration operations, obtain the target health care product representation data that can accurately reflect the health care needs and behavioral characteristics of the target users, providing important data support for the subsequent generation of accurate user portraits.

[0034] Step S400: Obtain any generated original portrait label, use the health care service behavior representation data and the target health care product representation data as label diffusion targets to reversely diffuse the original portrait label, obtain the target portrait label, and generate a target user portrait corresponding to the target user based on the target portrait label.

[0035] The computer system first obtains any generated original portrait labels. The original portrait labels can be preliminary labels based on some prior knowledge or randomly generated, which are used to describe some basic characteristics or potential attributes of the user. For example, in the smart health care scenario, the original portrait labels may include "age between 40-50 years old", "focus on health care", "have a certain sports foundation", etc. These labels provide an initial state for the subsequent reverse diffusion process. Then, the computer system reversely diffuses the original portrait labels with the health care service behavior representation data and the target health care product representation data as the label diffusion target. Label diffusion is a data-driven method that integrates specific data features (here, health care service behavior representation data and target health care product representation data) into the original portrait label so that the label can more accurately reflect the characteristics of the target user. Reverse diffusion is the process of evolving from a relatively rough initial state (original portrait label) to a more refined state (target portrait label) that is more in line with the actual situation of the target user during the diffusion process.

[0036] To illustrate with a specific example, suppose the original portrait label is "pay attention to health and wellness", the health and wellness service behavior representation data shows that the target user often participates in aerobic exercise (such as more than three times a week), and the target health and wellness product representation data shows that the user has purchased professional sports equipment (such as smart sports bracelets, high-quality sports shoes, etc.). When the computer system performs reverse diffusion, it will combine these behaviors and product features with the original portrait label, making the label gradually refined and accurate. For example, the original portrait label can gradually evolve into a more specific target portrait label such as "pay attention to health and wellness, prefer aerobic exercise, and pay attention to the quality of sports equipment."

[0037] In the process of realizing reverse diffusion, the computer system can adopt a variety of implementation methods. One feasible method is based on the idea of ​​generative adversarial network (GAN). Generative adversarial network consists of generator and discriminator. The generator tries to generate data as realistic as possible, while the discriminator tries to distinguish the generated data from the real data. In this step, the computer system can input the original portrait label, health care service behavior representation data and target health care product representation data into the generator. The generator gradually generates portrait labels that are closer to the real characteristics of the target user by learning the characteristics and distribution of these data.

[0038] Specifically, the computer system can use deep learning models to build generators and discriminators. For example, convolutional neural networks (CNN) or recurrent neural networks (RNN) are used. Taking CNN as an example, it has powerful feature extraction capabilities and can automatically learn local features and patterns in data. The computer system takes the original portrait label, health care service behavior representation data, and target health care product representation data as input, and performs feature extraction and dimensionality reduction through the multi-layer convolution layer and pooling layer of CNN. The extracted features are then input into the fully connected layer for further processing and conversion, and finally the target portrait label is generated.

[0039] Assume that the input of the generator is the original portrait label L0, the health care service behavior representation data B and the target health care product representation data P, and the output of the generator is the target portrait label L t The training process of the generator can be achieved by minimizing a certain loss function between the generated label and the true label, such as the cross entropy loss function, as follows: ; Among them, y i is a real label (it can be a real portrait label of the target user obtained through manual annotation or other reliable methods), is the target image label L generated by the generator t The i-th element in , n is the dimension of the label. By continuously adjusting the parameters of the generator, the value of the loss function is minimized, thereby improving the accuracy of the generated target image label.

[0040] After obtaining the target portrait label, the computer system generates a target user portrait corresponding to the target user based on the target portrait label. The target user portrait can be a comprehensive description containing a variety of information, including the user's basic characteristics, behavioral preferences, demand tendencies, etc. For example, based on the target portrait label "pay attention to health and wellness, prefer aerobic exercise, pay attention to the quality of sports equipment, and have a certain economic strength", the computer system can generate a target user portrait, describing the user as being between 40 and 50 years old, paying attention to physical health, doing aerobic exercise several times a week, liking to buy high-quality sports equipment, and having a good economic situation, and having high requirements for health care services.

[0041] The computer system can present the target user portrait in a variety of ways, such as describing the user's characteristics and preferences in detail in the form of a text report, or displaying the distribution of user characteristics in different aspects in a visual chart (such as a bar chart, pie chart, etc.). The target user portrait generated in this way can more accurately reflect the real situation of the target user, provide strong support for personalized recommendations and precision marketing of smart health care services, and improve the push conversion rate and user satisfaction of health care services.

[0042] In one implementation, in step S400, the original portrait label is reversely diffused using the health care service behavior representation data and the target health care product representation data as label diffusion targets to obtain the target portrait label, including: Step S410: loading the original portrait label, health care service behavior representation data, and target health care product representation data into the portrait label generation network, wherein the portrait label generation network includes a generative adversarial component, and the generative adversarial component includes a plurality of sequentially connected network layers; Step S420: For each network layer, the cross-modal association representation data obtained based on the health care service behavior representation data is added to the target health care product representation data and then mapped to obtain the execution result of the current network layer, and the target portrait label is obtained based on the execution result of the last network layer.

[0043] In step S410, the computer system first loads the original portrait label, health care service behavior representation data, and target health care product representation data into the portrait label generation network. The portrait label generation network is a deep learning network specially designed to generate target portrait labels. It contains a generative adversarial component, which is composed of multiple sequentially connected network layers.

[0044] For example, in a smart health care scenario, the original portrait label may be "pay attention to health and have a certain willingness to exercise", and the health care service behavior representation data may include information such as the number of times the user exercises per week, exercise duration, and exercise type. The target health care product representation data may include information such as the brand, price, and function of the sports equipment purchased by the user. The computer system loads this data into the portrait label generation network in a specific format and order. To ensure the validity and consistency of the data, the computer system can pre-process the data before loading, such as normalization and standardization. For numerical data, such as exercise duration, the formula To standardize, where x is the original data, is the mean of the data, is the standard deviation of the data, the standardized data It has the characteristics of mean 0 and standard deviation 1, which can improve the training effect and stability of the network.

[0045] In step S420, after loading the data into the portrait label generation network, the computer system performs specific calculations and processing for each network layer. Specifically, the cross-modal association representation data obtained based on the health care service behavior representation data and the target health care product representation data need to be added and mapped to obtain the execution result of the current network layer.

[0046] In order to obtain cross-modal association representation data, the computer system analyzes the association between the health care service behavior representation data and other modal data (such as text, images, etc., if any), and converts them into a unified representation form. For example, assuming that the types of exercise (such as running, swimming, etc.) in the health care service behavior representation data can be associated with related text descriptions (such as "aerobic exercise", "whole-body exercise", etc.), the computer system can convert these text descriptions into vector representations through the word vector model, and then merge them with the numerical representation of the exercise type to obtain cross-modal association representation data.

[0047] Next, the computer system adds the cross-modal association representation data to the target health care product representation data. For example, if the feature value representing the exercise intensity in the cross-modal association representation data is a, and the feature value representing the exercise intensity suitable for the sports equipment in the target health care product representation data is b, the result after addition is c=a+b.

[0048] Then, the computer system performs a mapping operation on the added result. The mapping operation can be implemented through the activation function in the neural network. The feasible activation functions include Sigmoid function, ReLU function, etc. Taking the ReLU function as an example, its formula is f(x)=max(0, x), that is, when the input value x is greater than 0, the output value is x; when the input value x is less than or equal to 0, the output value is 0. Through the mapping of the activation function, nonlinear factors can be introduced to enhance the expression ability of the network.

[0049] Assume that in a specific network layer, the result after addition is x, and the result after ReLU function mapping is y=f(x). This y is the execution result of the current network layer. The computer system calculates and processes each network layer in this way until it reaches the last network layer.

[0050] Finally, the target portrait label is obtained based on the execution result of the last network layer. For example, the execution result of the last network layer may be a multi-dimensional vector, and each dimension of the vector corresponds to a different label feature. The computer system can process this vector by setting a threshold or using a classification algorithm (such as a softmax classifier) ​​to convert it into a specific target portrait label. For example, for a dimension representing a user's sports preference, if its value is greater than a certain threshold, the corresponding target portrait label may be "love high-intensity sports"; if it is less than the threshold, it may be "prefer light sports".

[0051] In this way, the computer system can make full use of the health care service behavior characterization data and the target health care product characterization data, and through the calculation and processing of multiple network layers of the portrait label generation network, gradually generate more accurate and detailed target portrait labels, providing a key foundation for the final generation of the target user portrait.

[0052] In one implementation, in step S420, the cross-modal association representation data obtained based on the health care service behavior representation data is added to the target health care product representation data and then mapped to obtain the current network layer execution result, including: Step S421: performing internal attention focusing on the original image label or the execution result of the previous network layer to obtain first internal attention representation data; Step S422: generating a first search vector set according to the first internal attention representation data, generating a first content vector set and a first index vector set according to the health care service behavior representation data, performing cross-attention focusing processing according to the first search vector set, the first index vector set and the first content vector set, and obtaining first cross-attention focusing representation data; Step S423: Add the first cross-attention focus representation data and the target health care product representation data and then map them to obtain the execution result of the current network layer.

[0053] In step S421, when the computer system executes this step, it needs to perform an internal attention focusing operation on the original portrait label or the execution result of the previous network layer. The purpose of the internal attention mechanism is to enable the computer system to automatically learn the importance of different parts of the data and assign higher weights to the important parts, thereby extracting more representative and relevant feature information to form the first internal attention representation data.

[0054] For example, assuming that the original portrait label is "pay attention to health and wellness, and have certain exercise habits", this label can be represented as a vector form, such as ,in Represents the i-th element of the label vector. The computer system processes this vector through an internal attention mechanism. One possible implementation is to use a self-attention mechanism, which determines the importance weight of each element by calculating the correlation between each element in the vector and other elements.

[0055] Specifically, for the vector Each element in , the computer system will calculate it and other elements The correlation score between , such as the dot product operation, that is Then, these relevance scores are normalized by the softmax function to obtain the attention weights , the calculation formula of the softmax function is: ; where n is a vector Finally, the original vector is weighted and summed according to the attention weight to obtain the first internal attention representation data , the calculation formula is: ; In this way, the computer system can highlight the more important information parts in the original image label and provide more valuable data for subsequent processing. If the internal attention is focused on the execution result of the previous network layer, the operation process is similar, except that the input data becomes the result vector output by the previous network layer.

[0056] In step S422, after obtaining the first internal attention representation data, the computer system generates a first search vector set (query) based on the data. It can be transformed through a fully connected layer to obtain the first search vector set =[q1, q2,…, q m ], where q i represents the i-th search vector, and m is the number of search vectors.

[0057] At the same time, the computer system generates a first content vector set (value) and a first index vector set (key) based on the health care service behavior representation data. Assume that the health care service behavior representation data is ,in Represents the i-th health care service behavior characteristics. Similarly, through the full connection layer transformation, the first capacity vector set is obtained respectively and the first index vector set .

[0058] Next, the computer system performs cross-attention focusing processing based on the first search vector set, the first index vector set, and the first content vector set. Specifically, the search vector is first calculated. With index vector The correlation score between , such as the dot product operation, that is Then, these relevance scores are normalized by the softmax function to obtain the attention weights , the calculation formula is similar to the previous one: ; Finally, the content vectors are weighted and summed according to the attention weights to obtain the first cross-attention focus representation data , the calculation formula is: ; Through cross-attention focusing processing, the computer system can effectively associate and fuse the first internal attention representation data with the health care service behavior representation data, and extract the health care service behavior characteristic information related to the original portrait label.

[0059] For example, suppose the first internal attention representation data highlights the feature of "having certain exercise habits", while the health care service behavior representation data contains information such as the number of exercises per week and the duration of exercise. Through cross-attention focus processing, the computer system can find the relevant information such as the number of exercises and duration of exercise from the health care service behavior representation data based on the feature of "having certain exercise habits", and integrate it into the first cross-attention focus representation data.

[0060] In step S423, after obtaining the first cross-attention focus representation data, the computer system adds it to the target health care product representation data. Assume that the target health care product representation data is ,in Represents the characteristics of the i-th target health care product. The first cross-attention focus represents the data Characterization data of target health care products Add and get the result after addition .

[0061] The computer system then performs a mapping operation on the result of the addition to obtain the execution result of the current network layer. The mapping operation can be implemented by a fully connected layer and an activation function. For example, Input into a fully connected layer, the weight matrix of the fully connected layer is , the bias term is , then the output of the fully connected layer is .

[0062] Next, an activation function is used to perform a nonlinear transformation on the output of the fully connected layer. A feasible activation function is the ReLU function, whose calculation formula is f(x)=max(0, x). Activate through the ReLU function to get the execution result of the current network layer .

[0063] Through such operations, the computer system effectively integrates and maps the first cross-attention focus representation data with the target health care product representation data, and obtains the execution result of the current network layer. This result contains comprehensive information of the original portrait label, health care service behavior representation data, and target health care product representation data, laying the foundation for obtaining the target portrait label based on the execution result of the final network layer.

[0064] In one implementation, in step S410, before the original portrait label, health care service behavior representation data, and target health care product representation data are loaded into the portrait label generation network, the user portrait generation method applied to smart health care also includes: Step S401: obtaining an example health care service behavior sequence set and an example portrait tag of the example health care service behavior sequence set, and extracting example health care service behavior representation data and example health care product representation data of the example health care service behavior sequence set; Step S402: by generating a portrait label network, taking the example health care service behavior representation data and the example health care product representation data as label diffusion targets, reversely diffusing the example portrait labels after adding prior disturbance information, and obtaining example disturbed image labels; Step S403: Determine a training cost based on the error between the example perturbed image label and the prior perturbation information, and train the image label generation network according to the training cost.

[0065] In step S401, the computer system first obtains a set of example health care service behavior sequences and corresponding example portrait labels. The set of example health care service behavior sequences is a representative part of data selected from a large amount of historical health care service data, which covers the behavior records of different users in various health care service scenarios. For example, in a smart health care community, the computer system can select a part of users' health care service behavior sequences from the user data of the past year as example data. These behavior sequences may include users' daily exercise records (such as exercise type, exercise duration, exercise intensity, etc.), health monitoring data (such as blood pressure, heart rate, sleep quality, etc.), and records of health care activities participated in (such as attending health lectures, rehabilitation training courses, etc.).

[0066] The example portrait label is a description of the user characteristics corresponding to each example health care service behavior sequence. These labels are determined by manual annotation or other reliable methods. For example, for a user who often participates in aerobic exercise, regularly monitors health indicators, and pays attention to nutritional matching, his example portrait label may be "love sports, pay attention to health management, and pay attention to dietary nutrition."

[0067] After obtaining the set of example health care service behavior sequences and example portrait labels, the computer system extracts example health care service behavior representation data and example health care product representation data. The process of extracting example health care service behavior representation data is similar to the method of extracting health care service behavior representation data in step S100, mainly including operations such as data preprocessing, feature extraction and feature representation. For example, for the user's exercise duration data, the computer system can first perform data cleaning, remove outliers and missing values, and then calculate its mean, standard deviation and other statistical features as part of the representation data; for exercise type data, it can be converted into a vector representation, such as representing "running" as [1, 0, 0], "swimming" as [0, 1, 0], etc. The process of extracting example health care product representation data is similar to the method of extracting health care product representation data in step S300. The computer system analyzes the various attributes and characteristics of health care products associated with the example health care service behavior sequence and converts them into a suitable representation form. For example, for a smart sports bracelet purchased by a user, the computer system can extract its brand, function, price and other features, and then standardize or vectorize these features to obtain sample health care product characterization data.

[0068] In step S402, after obtaining the example health care service behavior characterization data and the example health care product characterization data, the computer system uses them as label diffusion targets and reversely diffuses the example portrait labels after adding prior disturbance information through the portrait label generation network.

[0069] Prior perturbation information is some random noise or specific perturbation pattern introduced to increase the diversity of data and the generalization ability of the model. For example, the computer system can add some random noise that follows a Gaussian distribution to the example image label, that is, for the example image label vector , the label vector after adding the prior perturbation information ,in It follows a Gaussian distribution A random noise vector, is the standard deviation of the noise.

[0070] After receiving the example portrait labels after adding prior disturbance information, as well as the example health care service behavior representation data and the example health care product representation data, the portrait label generation network will perform a reverse diffusion process. The purpose of reverse diffusion is to gradually recover from the state after adding disturbance to a state closer to the real example portrait label, so as to learn the potential characteristics and laws in the data. During the reverse diffusion process, the portrait label generation network will gradually adjust the feature representation of the label through calculations and processing of multiple network layers. For example, in a portrait label generation network based on a generative adversarial network (GAN), the generator will try to generate labels similar to the real example portrait labels based on the example health care service behavior representation data and the example health care product representation data. Specifically, the generator starts with the example portrait label after adding prior disturbance information, and calculates through a series of neural network layers (such as convolutional layers, fully connected layers, etc.), continuously updating the feature representation of the label so that it gradually approaches the real example portrait label.

[0071] In step S403, after obtaining the example perturbed image label, the computer system determines the training cost based on the error between the example perturbed image label and the prior perturbation information, so as to train the image label generation network.

[0072] There are many ways to calculate the error, such as Mean Squared Error (MSE). Assume that the example perturbed image label is , the prior disturbance information is , then the calculation formula of mean square error is: ; Where n is the dimension of the label vector, is the example scrambled image label vector The i-th element of is the prior perturbation information vector The ith element of .

[0073] The training cost is usually defined based on the error. For example, the training cost can be defined as the mean square error plus some regularization terms to prevent the model from overfitting. The regularization term can be a weight decay term, which is used to limit the size of the model parameters. The calculation formula is: ; in, is the regularization coefficient, It is the jth parameter in the image label generation network.

[0074] The process of training the image label generation network based on the training cost is usually implemented using an optimization algorithm, such as Stochastic Gradient Descent (SGD) and its variants (such as Adagrad, Adadelta, Adam, etc.). Taking the Adam optimization algorithm as an example, it will update the parameters of the image label generation network based on the training cost. The update formula is as follows: ; ; ; ; ; in, are the first-order moment estimate and the second-order moment estimate, respectively. is the decay rate, is the gradient of the training cost with respect to the parameter w, is the learning rate, is a small constant used to prevent the denominator from being zero, and t is the number of iterations.

[0075] By continuously calculating the training cost and using the optimization algorithm to update the parameters of the portrait label generation network, the computer system can enable the network to gradually learn the relationship between the example health care service behavior representation data, the example health care product representation data and the example portrait label, thereby improving the accuracy and reliability of the network's generation of target portrait labels. For example, after multiple iterations of training, the portrait label generation network can generate target portrait labels that are more in line with the user's real characteristics based on the input health care service behavior representation data and target health care product representation data, providing strong support for the subsequent generation of accurate target user portraits.

[0076] In one implementation, in step S400, generating a target user portrait corresponding to the target user according to the target portrait tag includes: Step S430: performing reverse diffusion on the original image label without setting a label diffusion target to obtain a reference image label; Step S440: Integrate the target portrait label and the reference portrait label to obtain an integrated portrait label; Step S450: Generate a target user portrait corresponding to the target user based on the integrated portrait tag.

[0077] In step S430, when the computer system executes this step, it will perform a reverse diffusion operation on the original portrait label without setting a label diffusion target. The original portrait label is a preliminary label obtained in the previous step. It may be based on some prior knowledge or randomly generated. It has a certain degree of universality, but it is not accurate and personalized enough. Reverse diffusion is a process of gradually evolving from a relatively rough initial state to a more refined state that is more in line with the actual situation of the target user.

[0078] Without setting a label diffusion target, the computer system processes the original image labels mainly based on the structure of the image label generation network itself and the learned parameters. The image label generation network is usually composed of multiple network layers, which contain various neurons and connection weights. They have learned some characteristics and patterns of the data during the training process.

[0079] For example, suppose the original portrait label is "pay attention to health and wellness", the computer system inputs this label into the portrait label generation network. The network will start from this initial label and gradually refine and optimize the label through a series of calculations and transformations. In each layer of the network, neurons will process the input label information according to their connection weights, and may introduce some random noise or make adjustments based on the learned distribution. After being processed by multiple layers of the network, the original portrait label will gradually evolve to obtain a relatively more specific and richer reference portrait label. For example, after reverse diffusion, the reference portrait label may become "pay attention to health and wellness, tend to regulate the body through exercise and diet, and also pay some attention to mental health."

[0080] In step S440, after obtaining the target image tag and the reference image tag, the computer system integrates them to obtain a more comprehensive and accurate integrated image tag. The integration process needs to comprehensively consider the information contained in the two tags, avoid duplication and conflict of information, and make full use of the advantages of both.

[0081] One feasible way to integrate is through a ratio estimation network. The ratio estimation network is a neural network specially designed to estimate the ratio of the target image label and the reference image label. The computer system loads the target image label and the reference image label into the ratio estimation network, and the network learns and estimates the target selection ratios corresponding to the two labels based on their characteristics and information content.

[0082] For example, suppose the target portrait label is "love aerobic exercise, pay attention to the quality of sports equipment, and have a certain economic strength", and the reference portrait label is "pay attention to health and wellness, tend to regulate the body through exercise and diet, and also pay some attention to mental health". The computer system inputs these two labels into the ratio estimation network. After calculation, the network estimates that the target selection ratio of the target portrait label is 0.6, and the target selection ratio of the reference portrait label is 0.4.

[0083] Then, the computer system selects labels of corresponding proportions from the target image label and the reference image label according to the target selection ratio for integration. Specifically, for each feature in each label, the computer system performs weighted summation according to the target selection ratio. Assume that the feature vector in the target image label is represented as , the feature vector in the reference image label is expressed as , the target selection ratios are ( ), then the feature vector of the integrated image label The calculation formula is: ; Through this integration method, the computer system can fully combine the information of the target portrait label and the reference portrait label to obtain a more comprehensive and accurate integrated portrait label. For example, after integration, the integrated portrait label may become "love aerobic exercise, pay attention to the quality of sports equipment, have a certain economic strength, pay attention to health and wellness, tend to regulate the body through exercise and diet, and also pay some attention to mental health."

[0084] In step S450, after obtaining the integrated portrait tag, the computer system generates a target user portrait corresponding to the target user according to the integrated portrait tag. The target user portrait is a comprehensive and detailed description of the target user, which not only includes the user's basic characteristics and behavioral preferences, but also reflects the user's demand tendency and potential needs.

[0085] The computer system can generate the target user portrait in a variety of ways. One feasible way is to convert the integrated portrait tags into a natural language description to form a structured text report. For example, based on the integrated portrait tags "love aerobic exercise, pay attention to the quality of sports equipment, have a certain economic strength, pay attention to health and wellness, tend to regulate the body through exercise and diet, and also pay some attention to mental health", the computer system can generate the following target user portrait text report: "The target user is an individual who has a strong interest in health and wellness and has a certain level of financial strength. In terms of sports, they love aerobic exercise and pay great attention to the quality of sports equipment, and are willing to invest in high-quality sports equipment. In terms of health management, they not only pay attention to physical health, but also mental health, and tend to regulate their bodies through exercise and a reasonable diet to maintain a good life state." In addition to text reports, computer systems can also present target user portraits in a visual way. For example, by drawing bar graphs, pie charts and other charts to show the distribution of user characteristics in different aspects; or by using image generation technology to generate a virtual user image based on integrated portrait tags to vividly display the characteristics and preferences of the target user.

[0086] In order to make the target user portrait more vivid and specific, the computer system can also combine other relevant data and information, such as the user's historical consumption records, social behavior data, etc. For example, if the user has a history of purchasing high-end sports brand products, then the target user portrait can further emphasize his pursuit of quality and brand; if the user often shares health and wellness knowledge on social platforms, then his social activity and influence in the field of health and wellness can be reflected in the portrait.

[0087] Through the above steps, the computer system can start from the original portrait label, and after the processes of reverse diffusion, integration and portrait generation, it can finally generate an accurate and comprehensive target user portrait, provide strong support for personalized recommendations and precision marketing of smart health care services, and improve the push conversion rate and user satisfaction of health care services.

[0088] In one implementation, step S440, integrating the target portrait tag and the reference portrait tag to obtain an integrated portrait tag, includes: Step S441: Load the target image label and the reference image label into the ratio estimation network for estimation, and obtain the target selection ratios corresponding to the target image label and the reference image label respectively; Step S442: Select labels of corresponding proportions from the target portrait label and the reference portrait label according to the target selection ratio, and integrate them to obtain an integrated portrait label.

[0089] In step S441, when the computer system executes this step, it first loads the target portrait label and the reference portrait label into the ratio estimation network. The ratio estimation network is a neural network model specially designed to learn and predict the importance ratio of different labels. It analyzes and processes the input label data, learns the features and patterns therein, and thus estimates the proportion of each label in the final integrated portrait label.

[0090] The ratio estimation network usually consists of multiple neural network layers, including input layer, hidden layer and output layer. The input layer is responsible for receiving the data of target image label and reference image label, which are usually represented in the form of vectors.

[0091] In the hidden layer, neurons perform complex nonlinear transformations and feature extraction operations on the input label data. Viable hidden layer activation functions include ReLU (Rectified Linear Unit), Sigmoid, and Tanh. Taking the ReLU function as an example, its calculation formula is f(x)=max(0, x), which can effectively introduce nonlinear factors and enhance the network's expressive power. The hidden layer processes the input data through these activation functions and learns the potential features and patterns in the label data.

[0092] The output layer is responsible for outputting the target selection ratios corresponding to the target portrait label and the reference portrait label respectively. The target selection ratio is a value between 0 and 1, indicating the relative importance of the label in the integrated portrait label. Assume that the target selection ratio of the target portrait label is pT, the target selection ratio of the reference portrait label is pR, and pT+pR=1 is satisfied.

[0093] In order to train the proportion estimation network, the computer system uses a large amount of labeled data. These labeled data contain known target image labels, reference image labels, and their corresponding ideal target selection ratios. By inputting these labeled data into the proportion estimation network, using a suitable loss function (such as the mean square error loss function) to calculate the difference between the predicted ratio output by the network and the actual ratio, and using an optimization algorithm (such as the stochastic gradient descent algorithm) to update the parameters of the network, the network's prediction results gradually approach the actual ratio.

[0094] For example, suppose there is a set of labeled data, where the target portrait label is "love outdoor sports, pay attention to healthy eating", represented by the vector ; The reference portrait label is "pay attention to mental health and have regular physical examinations", represented as a vector ; The known ideal target selection ratio is The computer system will Input into the ratio estimation network, the network outputs the predicted target selection ratio after calculation Then, the loss value is calculated using the mean square error loss function, and the formula is: ; The computer system uses an optimization algorithm to update the parameters of the proportion estimation network based on the loss value, and continuously adjusts the weights and biases of the network so that the loss value gradually decreases and the predicted proportion becomes closer and closer to the actual proportion.

[0095] In step S442, after obtaining the target selection ratios corresponding to the target portrait label and the reference portrait label respectively, the computer system selects labels of corresponding ratios from the target portrait label and the reference portrait label according to these ratios for integration, thereby obtaining an integrated portrait label.

[0096] Specifically, for the target image label vector and the reference image label vector , and their corresponding target selection ratios , integrating the image label vector The calculation method is to perform a weighted sum of the characteristics of the target image label and the reference image label. The calculation formula is: ; For example, if the target profile label is "love aerobic exercise, pay attention to the quality of sports equipment, and have a certain economic strength", it is represented by the vector ; The reference portrait label is "pay attention to health and wellness, tend to regulate the body through exercise and diet, and also pay some attention to mental health", represented as a vector ; The target selection ratios are . Then the integrated image label vector The calculation process is as follows: ; This means that the integrated portrait label contains all the features of the target portrait label and the reference portrait label, and the importance of each feature is weighted according to the target selection ratio.

[0097] In actual applications, the features of the target image label and the reference image label may be of different data types, such as numerical, categorical, or textual. For numerical features, you can directly perform weighted summation according to the above formula; for categorical features, you can decide which category to select in the integrated image label based on the target selection ratio; for textual features, you can represent the text with word vectors and then perform weighted summation, or select part of the text for splicing based on the target selection ratio.

[0098] Through this integration method, the computer system can fully combine the information of the target portrait label and the reference portrait label, and reasonably integrate them according to their relative importance to obtain a more comprehensive and accurate integrated portrait label. This integrated portrait label comprehensively considers the specific target information reflected by the target portrait label and the general reference information provided by the reference portrait label, providing a richer and more reliable data foundation for generating accurate target user portraits.

[0099] In one implementation, step S200, performing behavior detection on a health care service behavior sequence in a health care service behavior sequence set to obtain an identification weight of a behavior identifier in the health care service behavior sequence, includes: Step S210: obtaining a behavior analysis annotation of a health care service behavior sequence in a health care service behavior sequence set, and extracting annotation indication representation data of the behavior analysis annotation, wherein the behavior analysis annotation is used to annotate a target behavior event that needs to be analyzed; Step S220: extracting behavior sequence representation data of the health care service behavior sequence, merging the behavior sequence representation data with the annotation indication representation data to obtain first merged representation data; Step S230: Perform behavior detection on the health care service behavior sequence according to the first combined representation data to obtain an identification weight of the behavior identifier.

[0100] In step S210, the computer system obtains the behavior analysis annotation of each health care service behavior sequence from the health care service behavior sequence set. Behavior analysis annotation is a kind of information that marks the health care service behavior sequence manually or through a specific algorithm, and its purpose is to clearly point out the target behavior events that need to be analyzed in detail. For example, in a health care service behavior sequence set that contains daily health care activity records of multiple elderly people, the behavior analysis annotation can mark the target behavior event of "whether the elderly have undergone rehabilitation training of a specified intensity within a specific time period." If an elderly person undergoes rehabilitation training of the specified intensity on Monday, Wednesday, and Friday mornings, the corresponding behavior analysis annotation will clearly record the target behavior events at these time points as "rehabilitation training of a specified intensity was performed."

[0101] Extracting the annotation indication representation data of the behavior analysis annotation is to convert the annotation information into a form that the computer system can understand and process. This may involve multiple implementation methods. For example, for the behavior analysis annotation in text form, the computer system can use natural language processing technology to perform pre-processing operations such as word segmentation and part-of-speech tagging on the annotated text. Assuming that the behavior analysis annotation is "the old man did light exercise on Tuesday", after word segmentation, we get "the old man" "did" "light exercise" on "Tuesday", and then we can use a word vector model (such as Word2Vec or GloVe) to map each word to a corresponding vector.

[0102] In addition, if the behavior analysis annotations contain some numerical information, such as the quantitative value of exercise intensity, exercise duration, etc., the computer system can directly standardize these values.

[0103] In step S220, the computer system extracts the behavior sequence representation data of the health care service behavior sequence. This process is similar to extracting the label indication representation data, but focuses more on the characteristics of the behavior sequence itself. For example, for a health care service behavior sequence containing exercise records, the computer system can extract the statistical characteristics of the behavior sequence, such as the exercise frequency (the number of days of exercise in a week), the mean and variance of the exercise duration, etc.

[0104] In addition to statistical features, computer systems can also extract temporal features of behavioral sequences. For example, Fourier transform is used to convert the time-varying sequence of exercise duration into the frequency domain to obtain periodic information. If it is found that the exercise duration of the elderly has obvious peaks on Mondays, Wednesdays, and Fridays, then the corresponding frequency components can be extracted through Fourier transform as part of the behavioral sequence representation data.

[0105] Next, the computer system merges the behavior sequence representation data with the label indication representation data to obtain the first merged representation data. This merging process is relatively complex and involves multiple steps. First, the label indication representation data is internally focused to obtain the second internal attention representation data. For example, for the word vector set in the label indication representation data, the computer system can use the self-attention mechanism to focus on important information. Assume that there are three word vectors in the label indication representation data. , the self-attention mechanism calculates the attention weights between each vector and other vectors. The relevance score is obtained by dot product operation, such as The relevance score , and then convert these scores into attention weights through the softmax function (The calculation is simplified here, and more vectors may be involved in practice.) Finally, the labeled indication representation data is weighted summed according to the attention weight to obtain the second internal attention representation data.

[0106] Generate a second search vector set based on the second internal attention representation data, generate a second content vector set and a second index vector set based on the behavior sequence representation data, perform cross-attention focusing processing based on these three sets, and obtain second cross-attention focusing representation data. For example, the second internal attention representation data is transformed into a second search vector set through a linear transformation. For features such as motion frequency and mean in the behavior sequence representation data, a second content vector set and a second index vector set are generated through different linear transformations. Then, by calculating the dot product between the search vector and the index vector, and then obtaining the attention weight through the softmax function, the content vector is weighted and summed to obtain the second cross-attention focusing representation data.

[0107] The second cross-attention focus representation data is mapped to obtain the mapped representation data. This can be achieved by a nonlinear function (such as ReLU function). Assuming that the second cross-attention focus representation data is a vector , after being processed by the ReLU function f(x)=max(0, x), the mapping representation data is obtained , where each element .

[0108] A third search vector set is generated based on the behavior sequence representation data, a third content vector set and a third index vector set are generated based on the mapping representation data, and cross-attention focusing processing is performed based on these three sets to obtain third cross-attention focusing representation data. Finally, a fourth search vector set is generated based on the mapping representation data, a fourth content vector set and a fourth index vector set are generated based on the third cross-attention focusing representation data, and cross-attention focusing processing is performed based on these three sets to obtain the first merged representation data. Through this series of complex operations, the computer system deeply integrates the behavior sequence representation data with the annotation indication representation data, providing more comprehensive and accurate data support for subsequent behavior detection.

[0109] In step S230, the computer system performs behavior detection on the health care service behavior sequence according to the first merged representation data to obtain the identification weight of the behavior identifier. The behavior detection process can be regarded as a classification problem, that is, judging whether the health care service behavior sequence contains a specific target behavior event. The computer system can use a variety of machine learning or deep learning models to perform behavior detection, such as support vector machine (SVM), decision tree, neural network, etc.

[0110] Taking neural network as an example, the computer system takes the first merged representation data as input and inputs it into a pre-trained neural network model. The neural network model may include multiple hidden layers and an output layer. In the hidden layer, neurons perform nonlinear transformations on the input data through activation functions (such as sigmoid function). ) converts the input signal into the output signal. Assume that the signal input to the neuron in the hidden layer is x, and after the activation function, the output is y= (x).

[0111] At the output layer, the neural network will output a probability value, indicating the possibility that the health care service behavior sequence contains the target behavior event. For example, an output value of 0.8 means that the behavior sequence has an 80% probability of containing the target behavior event. This probability value is the identification weight of the behavior identifier. The behavior identifier is a tool used to identify the target behavior event, and the identification weight reflects the probability of the target behavior event appearing in the current health care service behavior sequence.

[0112] Assuming that the target behavior event is "the elderly performed high-intensity exercise", after calculation by the neural network, the output identification weight is 0.7. This means that based on the current health care service behavior sequence and related annotation information, the computer system determines that the probability of the elderly performing high-intensity exercise is 70%. In this way, the computer system can accurately detect the health care service behavior sequence based on the first merged representation data, obtain the identification weight of the behavior identifier, and provide a key basis for subsequent data analysis and user portrait generation. These identification weights can help the computer system understand the user's behavior patterns more accurately, thereby providing more accurate support for smart health care services. For example, when recommending health care activities, products or services suitable for users, it can more accurately match the user's actual needs.

[0113] In one implementation, step S220, merging the behavior sequence representation data with the annotation indication representation data to obtain first merged representation data, includes: Step S221: performing internal attention focusing on the marked indication representation data to obtain second internal attention representation data; Step S222: generating a second search vector set according to the second internal attention representation data, generating a second content vector set and a second index vector set according to the behavior sequence representation data, performing cross-attention focusing processing according to the second search vector set, the second index vector set and the second content vector set, and obtaining second cross-attention focusing representation data; Step S223: Mapping the second cross-attention focus representation data to obtain mapping representation data; Step S224: generating a third search vector set according to the behavior sequence representation data, generating a third content vector set and a third index vector set according to the mapping representation data, performing cross-attention focusing processing according to the third search vector set, the third index vector set and the third content vector set, and obtaining third cross-attention focusing representation data; Step S225: Generate a fourth search vector set based on the mapping representation data, generate a fourth content vector set and a fourth index vector set based on the third cross-attention focus representation data, perform cross-attention focus processing based on the fourth search vector set, the fourth index vector set and the fourth content vector set to obtain the first merged representation data.

[0114] In step S221, when processing the data represented by the annotation indication, the computer system adopts an internal attention focusing mechanism, which aims to highlight the more important parts of the data so as to better capture the key information related to the target behavior event. The internal attention focusing calculates the correlation between the elements within the data and assigns different weights to different elements, so that the important elements receive more attention in subsequent processing.

[0115] For example, suppose the label indicates that the representation data is a vector set representing the user's motion behavior label , where each vector Represents a specific annotation feature, such as Indicates the type of exercise (such as running, swimming, etc.), Indicates the intensity of exercise (high, medium, low), Indicates the duration of exercise. Indicates the frequency of movement.

[0116] To calculate the internal attention weights, the computer system first calculates the correlation score between the vectors. One possible way to do this is to use a dot product operation, such as calculating The correlation score between , the formula is These relevance scores are then converted into attention weights through a softmax function , the calculation formula of the softmax function is: ; Where n is the number of vectors in the labeled indicator representation data.

[0117] Finally, the annotation indication representation data is weighted and summed according to the attention weight to obtain the second internal attention representation data , the calculation formula is: ; In this way, the computer system can assign different weights to different annotation features according to the internal correlation of the data, highlight the features that are more relevant to the target behavior events, and obtain the second internal attention representation data.

[0118] In step S222, after obtaining the second internal attention representation data, the computer system generates a second search vector set based on it. This is usually achieved through a linear transformation, such as using a weight matrix For the second internal attention representation data Perform multiplication to obtain the second search vector set , the formula is .

[0119] At the same time, the computer system generates a second content vector set and a second index vector set based on the behavior sequence representation data. Assume that the behavior sequence representation data is a vector set representing the user's motion behavior record , where each vector Represents a specific behavior sequence feature, such as Indicates the daily exercise duration record. Indicates the number of exercises per week. Indicates the change record of exercise intensity. Indicates the switching record of the motion type.

[0120] Generate a second content vector set and the second index vector set The method is similar to generating the second search vector set, using the weight matrix Behavior sequence representation data Perform a linear transformation to obtain .

[0121] Next, the computer system performs cross-attention focusing processing based on the second search vector set, the second index vector set, and the second content vector set. First, the second search vector is calculated. With the second index vector The correlation score between , we can also use the dot product operation, that is These relevance scores are then converted into attention weights through a softmax function , the calculation formula is: ; Where m is the number of vectors in the behavior sequence representation data.

[0122] Finally, the second content vector is weighted and summed according to the attention weight to obtain the second cross-attention focus representation data , the calculation formula is: ; Through cross-attention focus processing, the computer system can effectively associate and fuse the annotation indication representation data and the behavior sequence representation data, and extract the behavior sequence features related to the target behavior event.

[0123] In step S223, in order to further process and convert the second cross-attention focus representation data, the computer system performs a mapping operation on it. The mapping operation is usually implemented by a nonlinear function, and feasible nonlinear functions include ReLU function, Sigmoid function, etc. Here, the ReLU function is taken as an example, and its calculation formula is f(x)=max(0, x).

[0124] Assume that the second cross-attention focus representation data is vector , after being mapped by the ReLU function, the mapped representation data is obtained , where each element For example, if ,but ;if Through this mapping operation, the computer system can introduce nonlinear factors, enhance the expressiveness of data, and make the mapped representation data more suitable for subsequent processing.

[0125] In step S224, the computer system generates a third search vector set based on the behavior sequence representation data in a similar manner to the second search vector set, using a new weight matrix Behavior sequence representation data Perform linear transformation to obtain the third search vector set .

[0126] Representing data based on mapping Generate a third content vector set and the third index vector set , using the weight matrix Perform a linear transformation to obtain .

[0127] Then, calculate the third search vector With the third index vector The correlation score between , using the dot product operation we get . Then the relevance score is converted into attention weight through the softmax function , the calculation formula is: ; Finally, the third content vector is weighted and summed according to the attention weight to obtain the third cross-attention focus representation data , the calculation formula is: ; Through this step of cross-attention focus processing, the computer system further integrates the behavior sequence representation data and the mapping representation data to extract richer feature information.

[0128] In step S225, the computer system characterizes the data according to the mapping Generate the fourth search vector set , using the weight matrix Perform a linear transformation, that is .

[0129] Based on the third span of attention focus representation data Generate the fourth content vector set and the fourth index vector set , using the weight matrix Perform a linear transformation to obtain .

[0130] Next, calculate the fourth search vector With the fourth index vector The correlation score between , through dot product operation. Then the relevance score is converted into attention weight through softmax function , the calculation formula is: ; Finally, the fourth content vector is weighted and summed according to the attention weight to obtain the first merged representation data , the calculation formula is: ; Through this series of complex operations, the computer system deeply integrates and associates the behavior sequence representation data with the annotation indication representation data, and obtains the first merged representation data containing rich information, which provides a more comprehensive and accurate data basis for the subsequent behavior detection of the health care service behavior sequence based on this data, and helps to more accurately determine the identification weight of the behavior identifier, thereby providing strong support for the generation of user portraits for smart health care.

[0131] In one implementation scheme, behavior detection is performed on a health care service behavior sequence in a health care service behavior sequence set, and the process of obtaining the identification weight of the behavior identifier in the health care service behavior sequence is implemented through an identification weight determination network, wherein the identification weight determination network includes a behavior analysis annotation processing component, a behavior sequence data processing component, an identifier processing component, and an identifier generation component, the behavior analysis annotation processing component is used to process the behavior analysis annotation, the behavior sequence data processing component is used to process the input health care service behavior sequence, the identifier processing component is used to process the behavior analysis annotation as an identifier type annotation, and the identifier generation component generates a corresponding behavior identifier based on the features output by the behavior analysis annotation processing component, the behavior sequence data processing component, or the identifier processing component.

[0132] The acquired behavior analysis annotations are loaded into the behavior analysis annotation processing component or the identifier processing component to extract the corresponding annotation indication representation data based on the behavior analysis annotation processing component or the identifier processing component; the health care service behavior sequence is loaded into the behavior sequence data processing component to extract the corresponding behavior sequence representation data based on the behavior sequence data processing component; the behavior sequence representation data and the annotation indication representation data are cross-fused to obtain the first merged representation data; the first merged representation data is loaded into the identifier generation component to perform behavior detection on the health care service behavior sequence; a behavior identifier of the behavior corresponding to the annotation indication representation data is generated according to the detection result; and the behavior identifier and the identification weight of the behavior identifier are output.

[0133] The identification weight determination network consists of a behavior analysis and annotation processing component, a behavior sequence data processing component, an identifier processing component, and an identifier generation component. Each component has a clear division of labor and works together to determine the identification weight of the behavior identifier.

[0134] The main task of the behavior analysis annotation processing component is to process behavior analysis annotations. Behavior analysis annotations are descriptions or labels of specific behaviors in the health care service behavior sequence, which are used to indicate the target behavior events that need to be paid attention to and analyzed. For example, in a health care service behavior sequence, the behavior analysis annotation may be "perform aerobic exercise more than three times a week". This annotation clarifies that the target behavior to be analyzed is aerobic exercise, and there are specific requirements for the exercise frequency.

[0135] The computer system loads the acquired behavior analysis annotations into the behavior analysis annotation processing component. The component analyzes and processes these annotations and extracts the corresponding annotation indication representation data. This process may involve multiple implementation methods, such as natural language processing technology and feature encoding methods. For behavior analysis annotations in text form, the computer system can first perform preprocessing operations such as lexical analysis and syntactic analysis to decompose the annotated text into words or phrases and determine their parts of speech and grammatical structure. Then, these words or phrases are converted into vector representations through word vector models (such as Word2Vec, GloVe, etc.), and these vectors constitute part of the annotation indication representation data.

[0136] At the same time, for some specific attributes or numerical information in the annotation, such as the frequency information of "more than three times", feature encoding can be performed. For example, it can be converted into a vector representation using One-Hot Encoding. Assuming that the movement frequency is divided into three categories: "less than once", "one to three times", and "more than three times", then "more than three times" can be encoded as [0, 0, 1]. Finally, these different parts of the representation data are combined to obtain the complete annotation indication representation data.

[0137] The behavior sequence data processing component is responsible for processing the input health care service behavior sequence. The health care service behavior sequence contains the user's various health care service behavior records over a period of time, such as daily exercise duration, exercise type, health monitoring data, etc.

[0138] The computer system loads the health care service behavior sequence into the behavior sequence data processing component, which will perform feature extraction and representation learning on these data to obtain the corresponding behavior sequence representation data. Different processing methods will be used for different types of behavior data.

[0139] For numerical data, such as exercise duration and heart rate, the computer system can first clean the data to remove noise and outliers. Then, it can calculate some statistical features, such as mean, standard deviation, maximum value, minimum value, etc., as part of the behavior sequence representation data.

[0140] For categorical data, such as sports types (running, swimming, yoga, etc.), one-hot encoding or embedding can be used to convert them into vector form. For example, there are three types of sports: "running", "swimming" and "yoga". "Running" can be encoded as [1, 0, 0], "swimming" as [0, 1, 0], and "yoga" as [0, 0, 1]. Through these methods, the computer system converts the health care service behavior sequence into easy-to-process behavior sequence representation data.

[0141] The identifier processing component is mainly used to process the annotations of the behavior analysis as identifier types. The identifier type annotations refer to those annotation information directly related to the behavior identifier, such as the annotations of specific behavior patterns or behavior characteristics.

[0142] When the computer system loads the behavior analysis annotation into the identifier processing component, the component will parse and process the annotation according to the definition and rules of the identifier, extract the feature information related to the identifier, and convert it into a suitable representation form. For example, assuming that the behavior identifier is used to identify the behavior pattern of "high-intensity exercise for three consecutive days", the identifier processing component will find information related to high-intensity exercise and consecutive days from the behavior analysis annotation, and convert it into a corresponding vector representation as part of the annotation indication representation data.

[0143] The function of the identifier generation component is to generate a corresponding behavior identifier based on the features output by the behavior analysis annotation processing component, the behavior sequence data processing component or the identifier processing component. After obtaining the behavior sequence representation data and the annotation indication representation data, the computer system cross-fuses them to obtain the first merged representation data. The cross-fusion process can be implemented in a variety of ways, such as splicing, weighted summing or fusing the two representation data through an attention mechanism. Then, the computer system loads the first merged representation data into the identifier generation component to perform behavior detection on the health care service behavior sequence. The identifier generation component will determine whether the health care service behavior sequence conforms to a specific behavior pattern or feature based on the feature information in the first merged representation data, thereby generating a behavior identifier for the behavior corresponding to the annotation indication representation data. For example, if the features in the first merged representation data indicate that the user has performed high-intensity exercise for three consecutive days in the past week, the identifier generation component will generate a behavior identifier indicating "high-intensity exercise for three consecutive days".

[0144] Finally, the identifier generation component outputs the behavior identifier and the identifier weight of the behavior identifier. The calculation of the identifier weight can be based on a variety of methods, such as statistical analysis, machine learning models, or probabilistic models. A simple method is to determine the identifier weight based on the frequency of the behavior identifier in the health care service behavior sequence.

[0145] By using identification weights to determine the coordinated work of various components of the network, the computer system can effectively detect the behavior sequence of health care services, accurately generate behavior identifiers and determine their identification weights, and provide important basic data for subsequent user portrait generation. These data can help the computer system to more deeply understand the user's health care service behavior characteristics and needs, thereby providing users with more accurate and personalized smart health care services.

[0146] In one implementation, in step S400, the original portrait label is reversely diffused using the health care service behavior representation data and the target health care product representation data as label diffusion targets to obtain the target portrait label, including: Step S400A: merging the health care service behavior representation data with the target health care product representation data to obtain second merged representation data; Step S400B: using the second combined representation data as a label diffusion target to perform reverse diffusion on the original portrait label to obtain a target portrait label.

[0147] In step S400A, when the computer system executes this step, it needs to merge the health care service behavior representation data with the target health care product representation data to obtain more comprehensive and integrated second merged representation data. This process involves the effective integration of two different types of but interrelated data to provide richer information for the subsequent reverse diffusion process.

[0148] Health care service behavior characterization data is an abstract representation of the target user's behavior characteristics in health care services. It contains the user's behavior records and feature extraction results in various health care service scenarios. For example, suppose in a smart health care community, the computer system collects the target user's health care service behavior data over a period of time, including the number of exercises per week, the duration of each exercise, the type of exercise (such as running, yoga, etc.), the frequency of participation in health lectures, etc.

[0149] The target health care product representation data is the characteristic representation of health care products associated with the health care service behaviors of the target users. These health care products may include sports equipment, health monitoring equipment, nutritional supplements, etc. For example, for sports enthusiasts, their related health care products may include smart sports bracelets, professional sports shoes, etc.

[0150] There are many ways for computer systems to merge these two representations of data, such as concatenation, weighted summation, and attention-based mechanisms.

[0151] In step S400B, after obtaining the second merged representation data, the computer system uses it as a label diffusion target and performs a reverse diffusion operation on the original portrait label to obtain a target portrait label that is more accurate and more in line with the characteristics of the target user.

[0152] The original portrait label is a preliminary description of the target user generated in the initial stage. It may be based on some prior knowledge or simple statistical information and is relatively rough and general. For example, the original portrait label may be "pay attention to health and wellness, and have a certain foundation in exercise." Backward diffusion is a process of evolving from a relatively simple and rough state to a more complex and sophisticated state. In this process, the rich information contained in the second merged representation data will be gradually integrated into the original portrait label, making it continuously refined and optimized.

[0153] The implementation of back diffusion in computer systems is usually based on generative models in deep learning, such as generative adversarial networks (GANs) or variational autoencoders (VAEs). The following is an example of a generative adversarial network: a generative adversarial network consists of a generator and a discriminator. The task of the generator is to generate a target image label that is as close to the real data distribution as possible based on the input second merged representation data, while the discriminator is responsible for determining whether the generated target image label is real or generated by the generator.

[0154] In the reverse diffusion process, the computer system first inputs the original portrait label and the second merged representation data into the generator. The generator will gradually adjust and optimize the original portrait label based on the information in the second merged representation data. For example, the generator can refine the original portrait label "pay attention to health and wellness, have a certain sports foundation" into "pay attention to health and wellness, do aerobic exercise more than 3 times a week, and prefer to use sports bracelets with heart rate monitoring functions" based on the specific information about the user's exercise behavior and the use of health care products in the second merged representation data.

[0155] The specific implementation of the generator can be a deep neural network containing multiple neural network layers. In each layer, the neurons will perform nonlinear transformations on the input data, and introduce nonlinear factors through activation functions (such as ReLU, Sigmoid, etc.) to enhance the network's expressiveness. For example, in a hidden layer, the input of the neuron is x, and after being processed by the activation function ReLU, the output is y=max(0, x).

[0156] After the generator generates the target image label, the discriminator will judge it. The discriminator is also a neural network. It learns the difference between the real image label and the image label generated by the generator, and outputs a judgment result, indicating the probability that the input image label is real. For example, the discriminator can output a value between 0 and 1, where 0 indicates that it is judged as a generated image label and 1 indicates that it is judged as a real image label.

[0157] The computer system continuously trains the generator and the discriminator to make the target image label generated by the generator closer and closer to the real image label, thereby completing the reverse diffusion process. The training process usually adopts the adversarial training method, that is, optimizing the network parameters by minimizing the game loss between the generator and the discriminator. Specifically, the goal of the generator is to maximize the probability that the discriminator will identify the target image label it generates as real, while the goal of the discriminator is to correctly distinguish the real image label from the generated image label. Through this adversarial training, the performance of the generator and the discriminator will continue to improve, and finally obtain accurate target image labels.

[0158] For example, after multiple iterations of training, the target portrait label generated by the generator based on the second merged representation data can be further refined into "pay attention to health and wellness, do aerobic exercise more than 3 times a week, each exercise lasts between 40 and 60 minutes, prefer to use sports bracelets with heart rate monitoring and sleep monitoring functions priced between 200 and 300 yuan, pay attention to nutritional matching, and regularly attend health lectures." Such target portrait labels can more accurately reflect the characteristics and needs of target users, and provide strong support for personalized recommendations and precision marketing of smart health services.

[0159] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0160] The memory 1002 stores computer programs that can be run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001. It can also cache data to be processed or processed by the processor 1001 and various modules in the computer system 1000 (for example, image data, audio data, voice communication data, and video communication data). This can be achieved through flash memory (FLASH) or random access memory (Random Access Memory, RAM).

[0161] When the processor 1001 executes the program, the steps of any of the above-mentioned user portrait generation methods for smart health care are implemented. The processor 1001 generally controls the overall operation of the computer system 1000.

[0162] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A user portrait generation method applied to smart health care, characterized in that: include: Acquire a health care service behavior sequence set collected for a target user, and extract health care service behavior representation data of the health care service behavior sequence set; Performing behavior detection on the health care service behavior sequence in the health care service behavior sequence set, and obtaining an identification weight of a behavior identifier in the health care service behavior sequence, wherein the behavior identifier is used to identify a target behavior event in the health care service behavior sequence, and the identification weight is a parameter for evaluating the predicted probability of the behavior identifier, and is used to indicate the probability of the target behavior event identified by the behavior identifier; Extracting first health care product representation data of the health care product associated with the target behavior event, integrating the first health care product representation data in the same health care service behavior sequence by using the identification weight to obtain second health care product representation data of the health care service behavior sequence, and integrating the second health care product representation data of multiple health care service behavior sequences to obtain target health care product representation data; Obtain any generated original portrait label, use the health care service behavior representation data and the target health care product representation data as label diffusion targets to reversely diffuse the original portrait label, obtain the target portrait label, and generate a target user portrait corresponding to the target user based on the target portrait label.

2. The user portrait generation method for smart health care according to claim 1 is characterized in that: The reverse diffusion of the original portrait label using the health care service behavior representation data and the target health care product representation data as label diffusion targets to obtain the target portrait label includes: Loading the original portrait label, the health care service behavior representation data, and the target health care product representation data into a portrait label generation network, wherein the portrait label generation network includes a generative adversarial component, and the generative adversarial component includes a plurality of sequentially connected network layers; For each of the network layers, the cross-modal association representation data obtained based on the health care service behavior representation data is added to the target health care product representation data and then mapped to obtain the execution result of the current network layer, and the target portrait label is obtained based on the execution result of the last network layer.

3. The user portrait generation method for smart health care according to claim 2 is characterized in that: The step of adding the cross-modal association representation data obtained according to the health care service behavior representation data and the target health care product representation data and then mapping them to obtain the current execution result of the network layer includes: Performing internal attention focusing on the original image label or the execution result of the previous network layer to obtain first internal attention representation data; Generate a first search vector set according to the first internal attention representation data, generate a first content vector set and a first index vector set according to the health care service behavior representation data, perform cross-attention focusing processing according to the first search vector set, the first index vector set, and the first content vector set, and obtain first cross-attention focusing representation data; The first cross-attention focus representation data and the target health care product representation data are added and mapped to obtain the current execution result of the network layer.

4. The user portrait generation method for smart health care according to claim 2 is characterized in that: Before the original portrait label, the health care service behavior representation data, and the target health care product representation data are loaded into the portrait label generation network, the user portrait generation method for smart health care further includes: Obtaining an example health care service behavior sequence set and an example portrait tag of the example health care service behavior sequence set, and extracting example health care service behavior representation data and example health care product representation data of the example health care service behavior sequence set; By using the portrait label generation network, the example health care service behavior representation data and the example health care product representation data are used as label diffusion targets, and the example portrait labels after adding prior disturbance information are reversely diffused to obtain example disturbed image labels; A training cost is determined based on an error between the example perturbed image label and the prior perturbation information, and the image label generation network is trained according to the training cost.

5. The user portrait generation method for smart health care according to claim 1 is characterized in that: Generating a target user portrait corresponding to the target user according to the target portrait tag includes: Without setting a label diffusion target, reversely diffuse the original image label to obtain a reference image label; Integrate the target image label and the reference image label to obtain an integrated image label; A target user portrait corresponding to the target user is generated based on the integrated portrait tag.

6. The user portrait generation method for smart health care according to claim 5 is characterized in that: The step of integrating the target image label and the reference image label to obtain an integrated image label includes: Loading the target image label and the reference image label into a ratio estimation network for estimation, and obtaining the target selection ratios corresponding to the target image label and the reference image label respectively; According to the target selection ratio, labels of corresponding ratio are selected from the target portrait label and the reference portrait label for integration to obtain an integrated portrait label.

7. The user portrait generation method for smart health care according to claim 1 is characterized in that: The performing behavior detection on the health care service behavior sequence in the health care service behavior sequence set to obtain the identification weight of the behavior identifier in the health care service behavior sequence includes: Obtaining a behavior analysis annotation of a health care service behavior sequence in the health care service behavior sequence set, and extracting annotation indication representation data of the behavior analysis annotation, wherein the behavior analysis annotation is used to annotate the target behavior event that needs to be analyzed; Extracting behavior sequence representation data of the health care service behavior sequence, merging the behavior sequence representation data with the annotation indication representation data to obtain first merged representation data; Performing behavior detection on the health care service behavior sequence according to the first combined representation data to obtain an identification weight of a behavior identifier; The step of merging the behavior sequence representation data with the annotation indication representation data to obtain first merged representation data includes: Performing internal attention focusing on the marked indication representation data to obtain second internal attention representation data; Generate a second search vector set according to the second internal attention representation data, generate a second content vector set and a second index vector set according to the behavior sequence representation data, perform cross-attention focusing processing according to the second search vector set, the second index vector set and the second content vector set, and obtain second cross-attention focusing representation data; Mapping the second cross-attention focus representation data to obtain mapping representation data; Generate a third search vector set according to the behavior sequence representation data, generate a third content vector set and a third index vector set according to the mapping representation data, perform cross-attention focusing processing according to the third search vector set, the third index vector set and the third content vector set, and obtain third cross-attention focusing representation data; Generate a fourth search vector set based on the mapping representation data, generate a fourth content vector set and a fourth index vector set based on the third cross-attention focus representation data, perform cross-attention focus processing based on the fourth search vector set, the fourth index vector set and the fourth content vector set, and obtain first merged representation data.

8. The method for generating user portraits for smart health care according to claim 7, characterized in that: The process of performing behavior detection on the health care service behavior sequence in the health care service behavior sequence set and obtaining the identification weight of the behavior identifier in the health care service behavior sequence is realized by an identification weight determination network, wherein the identification weight determination network includes a behavior analysis annotation processing component, a behavior sequence data processing component, an identifier processing component, and an identifier generation component, the behavior analysis annotation processing component is used to process the behavior analysis annotation, the behavior sequence data processing component is used to process the input health care service behavior sequence, the identifier processing component is used to process the behavior analysis annotation, and the identifier generation component generates a corresponding behavior identifier according to the features output by the behavior analysis annotation processing component, the behavior sequence data processing component or the identifier processing component; The acquired behavior analysis annotations are loaded into the behavior analysis annotation processing component or the identifier processing component to extract the corresponding annotation indication representation data based on the behavior analysis annotation processing component or the identifier processing component; the health care service behavior sequence is loaded into the behavior sequence data processing component to extract the corresponding behavior sequence representation data based on the behavior sequence data processing component; the behavior sequence representation data and the annotation indication representation data are cross-fused to obtain the first merged representation data; the first merged representation data is loaded into the identifier generation component to perform behavior detection on the health care service behavior sequence; a behavior identifier of the behavior corresponding to the annotation indication representation data is generated according to the detection result; and the behavior identifier and the identification weight of the behavior identifier are output.

9. The user portrait generation method for smart health care according to claim 1 is characterized in that: The reverse diffusion of the original portrait label using the health care service behavior representation data and the target health care product representation data as label diffusion targets to obtain the target portrait label includes: Merging the health care service behavior representation data with the target health care product representation data to obtain second merged representation data; The second combined representation data is used as a label diffusion target to perform reverse diffusion on the original image label to obtain a target image label.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images