User portrait text generation method and system based on health management information push

Abstract

The embodiment of the invention discloses a user portrait text generation method and system based on health management information push, and the method comprises the steps: responding to push analysis authorization authentication information fed back by a target user, and obtaining a health management data set of the target user, the health management data set comprises a plurality of health behavior records, each health behavior record is composed of at least one physiological index monitoring data and corresponding behavior activity data; feature extraction processing is conducted on the health management data set, a health behavior feature set of the health behavior records is obtained, and the health behavior feature set comprises physiological state features and activity association features; based on a preset portrait generation model, performing feature fusion processing on the physiological state features and the activity association features to generate a user portrait text; and generating a health management information pushing strategy according to the user portrait text, and issuing the health management information pushing strategy to a health management service platform for directional pushing operation.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of big data push, and specifically to a method and system for generating user portrait text based on health management information push. Background Art

[0002] In the field of big data push, with the rapid development of information technology, vast amounts of data are being collected and analyzed to achieve more accurate information push. Big data push technology is widely used in various fields. It aims to analyze massive amounts of data to uncover potential user needs, thereby pushing information that meets their interests and needs to users, improving the efficiency and effectiveness of information dissemination.

[0003] However, most existing big data push technologies focus solely on a single dimension of data, such as a user's consumption behavior or browsing history, while ignoring other important aspects of user health and failing to fully understand the user. Furthermore, existing technologies lack effective feature extraction and fusion methods when processing complex and diverse data. This results in inaccurate user profiles and push strategies that fail to truly meet users' actual needs. This leads to a lack of relevance between pushed information and users, making it difficult to ensure the accuracy and effectiveness of push notifications and resulting in a waste of network resources. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for generating user portrait text based on health management information push, which is used to improve the problems of poor accuracy and comprehensiveness and waste of network resources in the existing technology in health management information push.

[0005] In the first aspect, an embodiment of the present invention provides a method for generating user portrait text based on health management information push, which is applied to a user portrait text generation system, the method comprising: in response to push analysis authorization authentication information fed back by a target user, obtaining a health management data set of the target user, the health management data set comprising a plurality of health behavior records, each health behavior record consisting of at least one physiological indicator monitoring data and corresponding behavioral activity data; performing feature extraction processing on the health management data set to obtain a health behavior feature set of the health behavior record, the health behavior feature set comprising physiological state features and activity association features; based on a preset portrait generation model, performing feature fusion processing on the physiological state features and the activity association features to generate a user portrait text; generating a health management information push strategy according to the user portrait text, and issuing the health management information push strategy to a health management service platform for targeted push operations.

[0006] In a second aspect, an embodiment of the present invention provides a user portrait text generation system, comprising: processor; a storage device having a computer program stored thereon, When the computer program is executed by the processor, the processor implements any of the methods for generating user portrait text based on health management information push.

[0007] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method for generating user portrait text based on health management information push are implemented.

[0008] It can be seen that the embodiments of the present invention have the following beneficial effects: first, by obtaining a health management data set containing multiple health behavior records, and accurately extracting a health behavior feature set based on feature extraction processing, the user's health behavior characteristics can be clearly presented; based on the feature fusion processing of the portrait generation model, different dimensional features can be integrated to generate accurate and personalized user portrait text; the health management information push strategy generated based on the portrait text fully considers the different health conditions and behavior patterns of users, making the push content more targeted; based on this, the health management information push strategy is sent to the health management service platform for targeted push, which can ensure that users receive health management information that truly meets their needs, improve the accuracy and effectiveness of information push, and reduce the waste of network resources caused by invalid push. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flowchart of a method for generating user portrait text based on health management information push provided by an embodiment of the present invention.

[0010] Figure 2 A schematic diagram of the basic structure of a user portrait text generation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0012] See also Figure 1 As shown in FIG, this figure is a flow chart of a method for generating a user portrait text based on health management information push provided by an embodiment of the present invention, which can be applied to a user portrait text generation system. Figure 1 As shown, the method includes steps 110 to 140.

[0013] Step 110: In response to the push analysis authorization authentication information fed back by the target user, a health management data set of the target user is obtained, where the health management data set includes multiple health behavior records, each health behavior record consisting of at least one physiological indicator monitoring data and corresponding behavioral activity data.

[0014] In an embodiment of the present invention, when a target user submits push analysis authorization authentication information on the interactive interface of a health management platform, such as the authorization section of an app, the user profile text generator immediately responds. For example, taking user X as an example, user X clicks the "Confirm Authorization for Push Analysis" option on their frequently used health management app to provide the system with authorization information. Upon receiving this information, the system collects the target user's health management data set from multiple different data sources. These data sources are diverse and include various wearable health monitoring devices, such as smartwatches, which can track a user's various physiological indicators in real time, such as blood oxygen saturation and respiratory rate. They also include behavioral activity data that users have independently entered on the health management platform, such as their daily dietary intake, including information on the types and portions of food consumed at each meal. For user X, their health management data set contains numerous health behavior records. In one record, the physiological indicator monitoring data includes the values ​​of physiological indicators continuously monitored over a period of time, while the corresponding behavioral activity data records the user's dietary behavior during that period, such as the type of staple and non-staple foods consumed and the types of nutrients consumed.

[0015] Step 120: performing feature extraction processing on the health management data set to obtain a health behavior feature set of the health behavior record, wherein the health behavior feature set includes physiological state features and activity association features.

[0016] It is understood that after the system obtains the health management data set, it will immediately perform feature extraction on this data. The core purpose is to accurately extract key and representative features from massive and complex data, thereby providing a basis for subsequent in-depth analysis and generating accurate user profiles. Optionally, the feature extraction process performed on the health management data set to obtain the health behavior feature set of the health behavior record includes: Step 121: performing time alignment processing on the physiological indicator monitoring data in the health behavior record to generate a time-aligned physiological indicator sequence.

[0017] Because physiological metric data collected from different data sources may have inconsistent time records, time series alignment is essential. For example, different brands of health monitoring devices may have varying accuracy and frequency in recording physiological metrics. For user X's health behavior records, the physiological metric data from a smartwatch may have precise timestamps, recorded in seconds; while some physiological metric data recorded by other devices may have more flexible timestamps, recorded in minutes. In such cases, the system applies time alignment algorithms, such as those based on timestamp matching and data interpolation, to normalize all physiological metric data to a consistent time interval. For example, if minutes are used as the consistent time interval standard, for data with mismatched timestamps, the system uses data from adjacent time points and a reasonable interpolation method to calculate the appropriate value corresponding to that minute. This generates a consistent physiological metric series, ensuring that each data point in the series accurately corresponds to a consistent time scale, providing an accurate temporal dimension for subsequent analysis.

[0018] Step 122: calling a preset physiological feature extraction model to perform periodic fluctuation analysis on the time-aligned physiological indicator sequence to obtain the physiological state characteristics of the health behavior record, wherein the physiological state characteristics include trend change characteristics and abnormal fluctuation characteristics of the physiological indicators.

[0019] Furthermore, the system will start a pre-trained physiological feature extraction model with targeted analysis capabilities. The main function of this model is to deeply analyze the physiological indicator sequence after time alignment, and to extract the hidden key features that can reflect the user's physiological state.

[0020] Preferably, the calling of a preset physiological feature extraction model to perform periodic fluctuation analysis on the time-aligned physiological indicator sequence to obtain the physiological state features of the health behavior record includes: Step 1221: input the time-aligned physiological indicator sequence into the temporal convolution module of the physiological feature extraction model, and extract the local trend features of the physiological indicator sequence through multi-layer convolution kernels.

[0021] The temporal convolution module of the physiological feature extraction model is equipped with multiple convolution kernels with different parameters and sizes. For example, consider user X's time-aligned physiological indicator sequence, which contains continuously monitored physiological indicator data over a set time period. Once this sequence is input into the temporal convolution module, each convolution kernel gradually slides across the sequence according to predefined rules. During this sliding process, each convolution kernel performs computational operations on the subsequence it covers. For example, a smaller convolution kernel might focus on changes in physiological indicators over a shorter time span. Through convolution operations, it extracts the local trend of changes in the physiological indicator within that time period, determining whether it is increasing, decreasing, or remaining relatively stable. Larger convolution kernels, on the other hand, focus on the dynamics of physiological indicators over longer time frames, capturing more macroscopic local trend features. These local trend features serve as important foundational data for subsequent analysis.

[0022] Step 1222: Input the local trend features into the attention allocation module of the physiological feature extraction model, and generate a temporal attention weight based on the correlation between the local trend features of different time windows.

[0023] The attention allocation module conducts a detailed analysis of the input local trend features. It compares the inherent correlations between local trend features extracted from different time windows. For example, it compares local trend features extracted from a short time window (e.g., 5 minutes) with those extracted from a longer time window (e.g., 30 minutes), analyzing their similarities and differences. This comparison enables the module to determine which time periods' local trend features are more critical for reflecting overall physiological status and which are less important. Based on these analysis results, the module generates corresponding temporal attention weights for local trend features in different time windows. These weights are used in subsequent weighted processing of local trend features to more accurately highlight the impact of important features on overall physiological status.

[0024] Step 1223: Perform weighted aggregation on the local trend features according to the time attention weight to obtain a global trend feature.

[0025] After generating the temporal attention weights, the system performs a weighted aggregation operation on the local trend features based on these weights. Specifically, for each local trend feature, the system multiplies it by the corresponding temporal attention weight, and then aggregates and merges all weighted local trend features. In this way, local trend features that are assigned higher weights (i.e., those that have a greater impact on the overall physiological state) will occupy a more important position in the aggregation results, while the influence of local trend features with lower weights will be relatively weakened. Ultimately, the result obtained after weighted aggregation is the global trend feature, which comprehensively reflects the overall change trend of physiological indicators over the entire time period, providing key information for a comprehensive understanding of the user's physiological state.

[0026] Step 1224: The global trend feature is combined with the statistical distribution feature of the time-series aligned physiological indicator sequence to generate the physiological state feature. The statistical distribution feature includes the mean, variance and skewness coefficient. The global trend feature includes a periodic trend component and a seasonal fluctuation component.

[0027] Optionally, the system will combine the previously obtained global trend features and the statistical distribution features of the time-aligned physiological indicator sequence. Statistical distribution features can describe the distribution characteristics of physiological indicator data from different angles. The mean can reflect the average level of the physiological indicator in the time period; the variance reflects the degree of dispersion of the data, that is, the fluctuation of the data around the mean; the skewness coefficient can describe the degree of asymmetry of the data distribution. The periodic trend component in the global trend feature reflects the periodic change pattern of the physiological indicators over time. For example, some physiological indicators may have daily cycles, weekly cycles and other change patterns; the seasonal fluctuation component takes into account the regular fluctuations that may occur in physiological indicators in different seasons. Combining these two types of features can comprehensively and multi-dimensionally characterize the physiological state characteristics of health behavior records, providing rich and comprehensive physiological information basis for the subsequent generation of accurate user portraits.

[0028] Step 123: Perform semantic parsing on the behavioral activity data in the health behavior record to generate a structured activity tag set.

[0029] The system performs in-depth semantic analysis of behavioral activity data in health behavior records, particularly data related to dietary management. For example, dietary activity data recorded by user X might include natural language descriptions, such as "I had a vegetable salad for lunch, paired with grilled chicken breast and whole-wheat bread." The system uses natural language processing techniques and a pre-built dietary semantic knowledge base to parse these descriptions. It identifies key information, such as food category (vegetable salad, grilled chicken breast, whole-wheat bread) and consumption time (lunch), and structures this information. In this way, complex natural language descriptions are converted into activity labels with clear semantics and structure, such as "lunch - vegetable salad - grilled chicken breast - whole-wheat bread." This ultimately generates a set of structured activity labels that clearly present the user's dietary behavior patterns and key information, facilitating further analysis of activity-related features.

[0030] Step 124: calling a preset behavior feature extraction model to perform context association analysis on the structured activity tag set to obtain activity association features of the health behavior record, wherein the activity association features include spatiotemporal distribution features and pattern repetition features of the behavior activity.

[0031] It can be understood that the system will enable a preset behavioral feature extraction model to analyze the structured activity tag set. The model is designed to mine various associated information hidden in the tag set, thereby extracting the key features of the behavioral activity.

[0032] Preferably, the calling of a preset behavior feature extraction model to perform context association analysis on the structured activity tag set to obtain the activity association features of the health behavior record includes: Step 1241: Map the structured activity tag set to a pre-generated behavior semantic space to generate an initial behavior semantic vector.

[0033] The pre-generated behavioral semantic space is a vector space constructed based on a large amount of dietary behavior data and related knowledge. It can map different dietary behavior labels to corresponding vector representations. For a set of structured activity labels generated by user X, the system projects each label into this behavioral semantic space according to the mapping rules. For example, the label "lunch - vegetable salad - grilled chicken breast - whole wheat bread" is converted into a corresponding initial behavioral semantic vector based on its position in the semantic space and its relationship to other labels. This vector contains semantic information about the dietary behavior, such as the association between food categories and the relationship between consumption time and food choice, providing a digital representation for subsequent analysis.

[0034] Step 1242: Call the spatial association module of the behavior feature extraction model to calculate the spatial distance features between the initial behavior semantic vectors, and generate a behavior association graph based on the spatial distance features.

[0035] The spatial association module of the behavioral feature extraction model analyzes the generated initial behavioral semantic vectors and calculates the spatial distance between different vectors. This distance reflects the semantic similarity or difference between different dietary behaviors. For example, for the vectors corresponding to the labels "lunch-vegetable salad-grilled chicken breast-whole wheat bread" and "dinner-vegetable soup-fried fish-brown rice," the similarity between the two dietary behaviors is determined by calculating their distance in the behavioral semantic space. A close distance indicates that the two dietary behaviors have certain similarities in terms of food selection and timing; otherwise, they differ significantly. Based on these spatial distance features, the system constructs a behavioral association graph. In this association graph, each node represents an initial behavioral semantic vector (i.e., a dietary behavior), and edges between nodes represent the association between two vectors. The edge weight is determined by the spatial distance. This behavioral association graph intuitively demonstrates the intrinsic connections between different dietary behaviors, providing a foundation for subsequent community segmentation and feature extraction.

[0036] Step 1243: performing community division processing on the behavior association graph to obtain multiple behavior community subgraphs, each behavior community subgraph corresponding to a behavior pattern category.

[0037] Among them, the system will perform community division operations on the constructed behavior association graph, and through the graph clustering algorithm, divide the closely connected nodes in the behavior association graph into a community to form multiple behavior community subgraphs. For example, the nodes corresponding to dietary behaviors with similar characteristics in terms of dietary structure, eating time, etc. will be divided into the same community. A community may represent a dietary pattern based on healthy light meals, and the dietary behaviors corresponding to the nodes may contain more vegetables, low-fat protein sources, etc.; another community may represent a traditional balanced dietary pattern, including a reasonable combination of various staple foods, meats, and vegetables. Each behavior community subgraph corresponds to a different behavior pattern category.

[0038] Step 1244: Extract the topological structure features and node distribution features of each behavioral community subgraph, and fuse the topological structure features with the node distribution features to generate the activity association features. The topological structure features include community density and centrality indicators, and the node distribution features include label distribution uniformity and time overlap.

[0039] For each behavioral community subgraph, the system extracts its topological structure and node distribution features. The community density in the topological structure reflects the closeness of connections between nodes in the community. A high community density indicates a strong correlation between dietary behaviors within the community. The centrality index can identify key nodes in the community. The dietary behaviors corresponding to these key nodes may have a significant impact on the behavioral patterns of the entire community. In terms of node distribution features, the label distribution uniformity can reflect the distribution of different types of dietary behavior labels within the community, such as whether a certain food category or consumption time dominates. The temporal overlap focuses on the degree of overlap of dietary behaviors within the community in the temporal dimension, such as whether the dietary behaviors of some communities are concentrated in a set time period. The system then integrates these topological structure and node distribution features, for example, through weighted calculation methods, to comprehensively process the values ​​of different features and ultimately generate activity association features. These features can comprehensively and meticulously describe the distribution patterns and pattern repetition characteristics of users' dietary behaviors in the spatiotemporal dimensions, providing important activity association information for the subsequent generation of user profiles.

[0040] Step 130: Based on a preset portrait generation model, perform feature fusion processing on the physiological state features and the activity-related features to generate a user portrait text.

[0041] The system will use a pre-set portrait generation model to deeply integrate the physiological state characteristics and activity-related characteristics extracted previously, thereby generating a user portrait text that can accurately describe the user's health management status and behavioral characteristics.

[0042] Optionally, the performing feature fusion processing on the physiological state feature and the activity-related feature based on a preset portrait generation model to generate a user portrait text includes: Step 131: Input the physiological state features and the activity-related features into the cross-modal alignment module of the portrait generation model to generate a feature alignment matrix.

[0043] Optionally, the cross-modal alignment module of the portrait generation model receives data from two different modalities: physiological state features and activity-related features. Since these two features come from different data sources and have different semantics and representations, the role of the cross-modal alignment module is to align and match them to a certain extent for better subsequent fusion. The module uses corresponding algorithms, such as calculating the similarity matrix between the two feature vectors, to generate a feature alignment matrix. This matrix can reflect the correspondence and similarity between the various dimensions of physiological state features and activity-related features, providing a basis for subsequent attention interaction and feature fusion. For example, for an indicator dimension reflecting cardiovascular health in the physiological state feature and a dimension related to exercise and diet in the activity-related feature, the feature alignment matrix can show the strength of the association between them, thereby helping the model better understand the potential connection between different features.

[0044] Step 132: Perform bidirectional attention interaction processing on the physiological state feature and the activity-related feature based on the feature alignment matrix to generate a fused feature vector.

[0045] Optionally, based on the generated feature alignment matrix, the system performs a bidirectional attention interaction operation on the physiological state features and activity-related features, which means that the model calculates the attention weights based on the physiological state features and activity-related features respectively to determine the importance of each feature in the fusion process.

[0046] Preferably, performing bidirectional attention interaction processing on the physiological state feature and the activity-related feature based on the feature alignment matrix to generate a fused feature vector includes: Step 1320: Using the physiological state feature as a query vector and the activity-related feature as a key vector and a value vector, calculate a first attention weight; perform weighted aggregation on the activity-related feature according to the first attention weight to obtain a first aggregated feature; using the activity-related feature as a query vector and the physiological state feature as a key vector and a value vector, calculate a second attention weight; perform weighted aggregation on the physiological state feature according to the second attention weight to obtain a second aggregated feature; concatenate the first aggregated feature and the second aggregated feature to generate the fused feature vector.

[0047] First, using physiological state features as query vectors and activity-related features as key and value vectors, the model calculates the first attention weight using a pre-defined calculation method, such as a feature alignment matrix and an attention mechanism formula. This weight reflects the importance of each component of the activity-related feature when focusing on the physiological state feature. Then, based on this weight, the activity-related features are weightedly aggregated. Specifically, the dimensions of the activity-related features are combined according to their weights to obtain the first aggregated feature. Conversely, using the activity-related features as query vectors and the physiological state features as key and value vectors, a similar calculation method is used to calculate the second attention weight. The physiological state features are then weightedly aggregated based on this weight to obtain the second aggregated feature. Finally, the first and second aggregated features are concatenated dimensionally, combining them in a specific order to form a new fused feature vector. This fused feature vector integrates the key information of both physiological state features and activity-related features. Through bidirectional attention interaction, it fully considers the mutual influence and importance of the two modal features, providing a high-quality feature representation for the subsequent generation of accurate user profile text.

[0048] Step 133: Input the fused feature vector into the text generation module of the portrait generation model, and expand the semantic paragraphs of the user portrait text layer by layer through the decoder. The semantic paragraphs include a health status summary, a behavior preference analysis, and a summary of intervention recommendations.

[0049] After the fused feature vector is input into the text generation module of the persona generation model, it uses its internal decoder to perform layer-by-layer expansion to generate text. The decoder gradually generates semantically meaningful text paragraphs based on the information contained in the fused feature vector. For example, in the health status summary, the decoder generates a description of the user's current overall health status based on information reflected by physiological status characteristics, such as "The user's recent physiological indicators show that cardiovascular function is stable, but blood sugar levels fluctuate during certain periods of time." In the behavioral preference analysis section, the decoder analyzes the user's dietary and other behavioral preferences based on activity-related features, such as "The user prefers foods rich in vegetables and high-quality protein and often has a richer meal at lunch." For the intervention recommendation summary, the decoder combines health status and behavioral preferences to provide targeted recommendations, such as "Given the user's blood sugar fluctuations and dietary preferences, it is recommended to appropriately increase dietary fiber intake and adjust the portion size and type of dinner food." Through this layer-by-layer expansion, a semantic paragraph containing multiple key components is generated.

[0050] Step 134: Perform a coherence check on the semantic paragraphs, delete redundant descriptions, adjust the order of the paragraphs, and generate the user portrait text.

[0051] In actual use, the generated semantic paragraphs may have some coherence issues or contain redundant information, necessitating coherence verification. The system uses natural language processing technology and pre-set coherence rules to check the logical relationships between paragraphs and the fluency of the sentences. For example, it determines whether there is a reasonable logical connection and content consistency between the health status summary, behavioral preference analysis, and intervention recommendation summary. For redundant descriptions, the system identifies and removes repetitive or unnecessary information to make the text more concise and clear. Paragraphs are also adjusted based on semantic rationality and logical order. For example, if part of the intervention recommendation section is found to be more closely related to the health status summary, the relevant content may be relocated to a more appropriate location. After the aforementioned coherence verification, redundant removal, and paragraph order adjustment, a logical, accurate, and fluent user profile is generated. This profile comprehensively and clearly describes the target user's health management status and behavioral characteristics, providing a strong basis for subsequent health management information push.

[0052] Step 140: Generate a health management information push strategy based on the user portrait text, and send the health management information push strategy to the health management service platform for targeted push operations.

[0053] Among them, based on the generated user portrait text, the system will further formulate and implement health management information push strategies to ensure that users are provided with accurate and effective health management advice and information.

[0054] In one example, the user portrait text is used to describe the health management demand characteristics and potential behavior preference characteristics of the target user, and the generation of a health management information push strategy based on the user portrait text includes: Step 141: Perform keyword extraction processing on the user portrait text according to the health management demand characteristics in the user portrait text to generate a health demand tag set.

[0055] The system analyzes the health management needs characteristics described in the user profile text. For example, if the user profile text mentions that the user's blood sugar fluctuates and has differentiated dietary preferences, the system will use a keyword extraction algorithm to extract key information from the text. The algorithm will identify words closely related to health management needs, such as "blood sugar fluctuations" and "dietary structure adjustment", and organize these keywords into a set of health needs tags. This set clearly summarizes the user's main health management needs and provides clear clues for subsequent matching and retrieval with the health knowledge base.

[0056] Step 142: Match and search the health requirement tag set with the pre-generated health knowledge base to obtain a candidate health management information set.

[0057] Among them, the pre-generated health knowledge base contains a large amount of organized and classified health knowledge and information, covering various health problems, solutions and related suggestions. The system will match and search the generated health need tag set with the health knowledge base. For example, taking the tag "blood sugar fluctuation" as an example, the system will search the knowledge base for all information related to blood sugar fluctuations, including the causes of blood sugar fluctuations, the impact on health, and the corresponding countermeasures. Through this matching retrieval, information related to the user's health need tags is filtered out from the knowledge base to form a candidate health management information set. The information in this set is potential and may be helpful for the user's health management.

[0058] Step 143: Sort the candidate health management information set according to the potential behavior preference features in the user portrait text to generate a priority queue.

[0059] Optionally, the user profile text also includes the user's potential behavioral preference characteristics, which are very important for determining the push priority of health management information. The system will sort the candidate health management information sets based on these characteristics.

[0060] Optionally, the sorting of the candidate health management information set according to the potential behavior preference features in the user portrait text to generate a priority queue includes: Step 1430: Extract the time period features describing the behavior time preference and the device type features describing the channel usage preference from the user portrait text; calculate the matching weight between the content theme of each piece of information in the candidate health management information set and the time period features; calculate the matching weight between the push channel adaptability of each piece of information in the candidate health management information set and the device type features; arrange the candidate health management information set in descending order according to the weighted sum of the matching weight and the matching weight to generate the priority queue.

[0061] Among them, the system first extracts key behavioral preference features from the user portrait text, such as the time period in which the user is more inclined to receive information, and which devices are usually used to obtain information. For example, the user portrait text shows that the user is more willing to check health information between 7pm and 9pm, and mainly receives information through mobile applications. For each piece of information in the candidate health management information set, the system calculates the matching weight between its content theme and the time period characteristics of the user's behavioral time preference. For example, if a piece of information is about blood sugar control after dinner, it has a high degree of match with the user's evening time period preference and may be given a higher matching weight. At the same time, the system calculates the fit weight between the push channel adaptability of each piece of information and the user's device type characteristics. If a piece of information is more suitable for push through a mobile application and has a high degree of fit with the user's device preference, it will also be given a higher fit weight.

[0062] For example, for the time period preference matching, the system will pre-set a time matching calculation rule to divide the user's time period preference into multiple sub-time periods, such as dividing the time period from 7pm to 9pm into two sub-time periods of 7pm-8pm and 8pm-9pm. When the content theme of a message matches the user's time period preference, the overlap ratio between the time involved in the message and the sub-time period is calculated. If the overlap ratio reaches 80% or above, the time period preference matching is considered high, and the message will be assigned a matching weight of 0.7-0.9; if the overlap ratio is between 50%-80%, the matching is medium, and a matching weight of 0.3-0.7 is assigned; if the overlap ratio is less than 50%, the matching is low, and a matching weight of 0-0.3 is assigned. For device preference compatibility, the system will determine it based on the match between the information push channel adaptability and the user device type characteristics. If the adaptability of the information push channel completely matches the characteristics of the user's device type, for example, the information is suitable for push through mobile applications and the user mainly uses mobile applications to receive information, then the device preference is considered to be highly compatible and a compatibility weight of 0.8-1.0 will be assigned. If there is a partial match, such as the information can also be pushed via SMS but the user prefers mobile applications, the compatibility is medium and a compatibility weight of 0.4-0.8 will be assigned. If there is basically no match, a compatibility weight of 0-0.4 will be assigned.

[0063] Finally, the system calculates the matching and compatibility weights based on pre-set weighting rules to arrive at a composite weight. This weight is then used to sort the candidate health management information sets in descending order, with higher-weighted information at the top, forming a priority queue. Information in this queue is ranked from highest to lowest in terms of its match with user behavior preferences, ensuring that the pushed information is more aligned with the user's habits and needs.

[0064] Step 144: Select a preset number of target health management information based on the priority queue, configure a push time strategy and a push channel strategy, and generate the health management information push strategy.

[0065] In detail, the system will select a corresponding number of target health management information from the priority queue based on the preset number. For example, 5 pieces of information are selected as push content by default. Then, according to the user's behavior time preference and device type preference, appropriate push time strategy and push channel strategy are configured for this information. For the users mentioned above who receive information through mobile applications between 7pm and 9pm, the system will arrange the selected target health management information between 7pm and 9pm and push it to the user through the mobile application. In the above way, a complete health management information push strategy is generated to ensure that the pushed information reaches the user at the right time and through the right channel.

[0066] In another possible embodiment, the sending of the health management information push policy to the health management service platform for targeted push operation includes: Step 145: Analyze the target health management information, push time strategy, and push channel strategy in the health management information push strategy.

[0067] Optionally, the system will conduct a detailed analysis of the generated health management information push strategy. It will identify the specific content of the target health management information, such as advice on blood sugar control and dietary guidance. It will also determine the push schedule, specifically the specific time schedule for each piece of information, and the push channel strategy, such as whether to push via a mobile app, SMS, or email. This analysis ensures that the health management service platform clearly understands the specific requirements of the push task.

[0068] Step 146: Generate a scheduled task instruction according to the push time strategy, and cache the target health management information in the push queue of the health management service platform.

[0069] Optionally, based on the push time strategy obtained through analysis, the system will generate a corresponding scheduled task instruction. For example, if the push time strategy stipulates that a piece of information about blood sugar monitoring after dinner should be pushed at 8 pm, the system will generate a scheduled task instruction to execute the push operation at 8 pm. At the same time, the system will temporarily store the target health management information in the push queue of the health management service platform, waiting to push it when the scheduled task instruction is triggered. The push queue plays the role of buffering and managing information, ensuring that the information is pushed in the predetermined order and time.

[0070] Step 147: When the execution time of the scheduled task instruction is reached, the interface service of the health management service platform is called, and the target health management information is sent to the terminal device of the target user according to the push channel strategy.

[0071] When the scheduled execution time arrives, the system calls the interface service provided by the health management service platform. For example, if the push channel strategy is to push through a mobile app, the system will call the mobile app's push interface to send the target health management information to the target user's mobile terminal. This interface service is responsible for accurately delivering the information to the user's device, ensuring that the user receives the health management information in a timely manner.

[0072] Step 148: Monitor the user interaction data of the target health management information, and update the potential behavior preference features in the user portrait text according to the user interaction data.

[0073] After pushing the target health management information, the system will closely monitor the user's interaction data with this information. For example, it records whether the user clicks on the pushed information, how long they stay on the information page, whether they forward it, and other operations. By analyzing these interaction data, the system can further understand the user's interests and behavior patterns. If it is found that users have a high number of clicks on a certain type of information about dietary nutrition and stay there for a long time, it means that the user is more concerned about this content. The system will update the potential behavioral preference characteristics in the user portrait text based on this data. For example, add a description in the user portrait text that the user is more interested in dietary nutrition information, so as to generate a more accurate health management information push strategy later.

[0074] In an independent embodiment, after the health management information push strategy is issued to the health management service platform for targeted push operation, it also includes: monitoring the real-time physiological indicator monitoring data of the target user after executing the health management information push strategy within a preset time period, and generating a feedback physiological indicator sequence; performing differential calculation on the feedback physiological indicator sequence and the corresponding physiological indicator monitoring data before pushing, to obtain an indicator change rate set, wherein the indicator change rate set includes the fluctuation amplitude and improvement direction of each physiological indicator; comparing the indicator change rate set with a preset intervention effect threshold, screening out abnormal physiological indicators that do not meet the standard and their associated health management information; based on the spatiotemporal distribution characteristics of the abnormal physiological indicators, adjusting the push content priority and push frequency in the health management information push strategy, generating an updated health management information push strategy and reissuing it.

[0075] Over a preset time period, the system continuously monitors the target user's real-time physiological indicator data after implementing a health management information push strategy. For example, after delivering health information about blood sugar management, the system continuously monitors the user's blood sugar level to generate a feedback physiological indicator sequence. This feedback sequence is then compared and analyzed with the corresponding physiological indicator monitoring data before the push. Through differential calculations, the changes in each physiological indicator are determined, forming a set of indicator change rates. This set clearly shows the magnitude of fluctuations in each indicator and whether it is improving or worsening. The system then compares this set of indicator change rates with a preset intervention effect threshold. If the magnitude of change in a physiological indicator does not meet the expected improvement standard, it is considered an abnormal physiological indicator that fails to meet the standard. The system then identifies the health management information associated with this abnormal physiological indicator, such as which information regarding diet control or exercise recommendations is relevant to the indicator. Based on the spatiotemporal distribution characteristics of abnormal physiological indicators, such as the time and frequency of occurrence, the system adjusts the health management information push strategy. If abnormal blood sugar fluctuations are frequent within a certain period of time, the push priority of information related to blood sugar management may be increased, and the push frequency may be increased. After the above adjustments, an updated health management information push strategy is generated and re-issued to the health management service platform to better meet the health management needs of users.

[0076] In an independent embodiment, after the health management information push strategy is sent to the health management service platform for targeted push operations, it also includes: collecting the target user's interactive operation log on the pushed health management information, extracting operation type characteristics and time response characteristics; mapping the operation type characteristics to a preset behavior feedback coding table, and generating a feedback intensity coefficient, the feedback intensity coefficient including click-through rate, stay time and secondary forwarding frequency; calculating the time overlap between the push information and the user's active period based on the time response characteristics, and generating a push effectiveness evaluation value in combination with the feedback intensity coefficient; performing backpropagation optimization on the feature fusion weights in the portrait generation model based on the push effectiveness evaluation value, and updating the parameter configuration of the behavior preference analysis module of the user portrait text.

[0077] The system collects logs of target users' interactions with pushed health management information. These logs record detailed information about the user's interactions with the information, such as when the user clicked on the information, how long they stayed on the information page, and whether they forwarded it a second time. The system extracts operation type features and time response features from these logs. Operation type features include different operations such as clicking, browsing, and forwarding; time response features record the specific time when users performed these operations. Next, the system maps the operation type features to a preset behavioral feedback coding table. For example, based on information such as click operations and dwell time, a feedback intensity coefficient is generated according to the rules of the coding table. The click-through rate can directly reflect the user's initial attention to the information, the dwell time reflects the depth of the user's interest in the information content, and the frequency of secondary forwarding indicates the information's potential for dissemination within the user's social circle.

[0078] For example, the system's preset behavioral feedback coding table sets different coefficient calculation rules for different operation type characteristics and corresponding time response characteristics. For click-through rate, the system calculates the ratio of the number of times users click on information to the total number of pushed information. If the click-through rate is 80% or above, a click-through rate coefficient of 0.8-1.0 is assigned; if it is between 50%-80%, a coefficient of 0.5-0.8 is assigned; if it is below 50%, a coefficient of 0-0.5 is assigned. For dwell time, the system divides dwell time into multiple intervals, such as less than 1 minute, 1-5 minutes, 5-10 minutes, and greater than 10 minutes, with different coefficients corresponding to each interval. For example, less than 1 minute corresponds to a coefficient of 0-0.2, 1-5 minutes corresponds to a coefficient of 0.2-0.5, 5-10 minutes corresponds to a coefficient of 0.5-0.8, and greater than 10 minutes corresponds to a coefficient of 0.8-1.0. For the secondary forwarding frequency, if the secondary forwarding frequency is 0, a coefficient of 0 is assigned; if it is forwarded 1-3 times, a coefficient of 0.2-0.5 is assigned; if it is forwarded 3-5 times, a coefficient of 0.5-0.8 is assigned; if it is forwarded more than 5 times, a coefficient of 0.8-1.0 is assigned. Finally, the system weights the click-through rate coefficient, the dwell time coefficient, and the secondary forwarding frequency coefficient according to the set weights (for example, the click-through rate coefficient has a weight of 0.4, the dwell time coefficient has a weight of 0.4, and the secondary forwarding frequency coefficient has a weight of 0.2) and adds them together to obtain the feedback strength coefficient.

[0079] At the same time, the system will calculate the time overlap between the push information and the user's active period based on the time response characteristics. If the user frequently interacts with the push information during his active period, it means that the push time is relatively appropriate. Combining the time overlap with the feedback intensity coefficient, the system will calculate a push effectiveness evaluation value, which can comprehensively measure the effect of health management information push. Based on this evaluation value, the system will use the back propagation algorithm to adjust the feature fusion weights in the portrait generation model. For example, if the push effectiveness evaluation value is low, it means that the user portrait and push strategy generated by the current model may be insufficient. The system will adjust the feature fusion weights through back propagation optimization, and then update the parameter configuration of the behavior preference analysis module in the user portrait text, so that the model can generate user portraits and push strategies that better meet user needs.

[0080] For example, the system will perform a weighted fusion of the time overlap and the feedback strength coefficient to calculate the push effectiveness evaluation value. First, the system will calculate the ratio of the overlap between the sending time of the push information and the user's active period to the total length of the user's active period to obtain the time overlap. Then, the time overlap and the feedback strength coefficient are weighted and summed according to the preset weights. For example, the weight of the time overlap is 0.3, and the weight of the feedback strength coefficient is 0.7. The push effectiveness evaluation value = time overlap × 0.3 + feedback strength coefficient × 0.7. In this way, the effect of health management information push is comprehensively measured.

[0081] In an independent embodiment, after the health management information push strategy is sent to the health management service platform for targeted push operations, it also includes: obtaining the health behavior records of the target user added after the push, performing anomaly detection on the physiological indicator monitoring data in the newly added records, and generating purified time series data after removing noise data; performing sliding window matching on the purified time series data with the historical health behavior feature set, extracting data distribution offset and pattern variation features; adjusting the statistical distribution features of the health behavior feature set according to the distribution offset, and using the pattern variation features to expand the pattern repetition features in the activity association features; inputting the updated health behavior feature set into the portrait generation model for training, and generating optimized user portrait text and associated push strategies.

[0082] The system captures new health behavior records generated by target users after health management information is pushed. These records contain new physiological indicator monitoring data and behavioral activity data. For the physiological indicator monitoring data in these new records, the system uses an outlier detection algorithm to identify data points that significantly deviate from the normal range, treating them as noise data and removing them.

[0083] For example, when the system detects outliers in newly recorded physiological indicator monitoring data, it first performs preliminary preprocessing on the data, including removing data with obvious errors (such as extremely large or small values ​​outside the normal range of physiological indicators). The system then compares and analyzes the newly recorded data with historical data, calculating the deviation of each data point from the historical mean. The system also considers the time series nature of the data and observes the data's changing trends. If the deviation of a data point exceeds a pre-set standard deviation multiple (e.g., 3 standard deviations), and the trend of that data point is significantly inconsistent with the overall trend of the historical data, the system will identify it as an outlier. For example, for a user's heart rate monitoring data, if the historical heart rate mean is 70 beats / minute and the standard deviation is 5 beats / minute, if a newly recorded heart rate data point reaches 90 beats / minute and the heart rate data before and after this point shows a sudden increase instead of a gradual increase, then this data point will be identified as an outlier.

[0084] Furthermore, the system can use the isolation forest algorithm, an unsupervised learning algorithm based on decision trees. The specific monitoring process is as follows: First, the system randomly extracts a portion of samples from the newly added physiological indicator monitoring data as a training set and uses these samples to construct multiple decision trees. Each decision tree starts at the root node, randomly selects a feature and a split point, and divides the data space into two subspaces. This process is repeated until every sample is isolated. During the decision tree construction process, outliers are isolated more quickly due to their significant distribution difference from normal data, meaning that their path length in the decision tree is shorter. For each data point in the newly added data, the system inputs it into the constructed decision tree and calculates its path length within the decision tree. The system sets an outlier threshold. When the path length of a data point is less than this threshold, it is identified as an outlier, treated as noise data, and removed. For example, for the user's blood pressure monitoring data, the system constructs 50 decision trees and calculates the path length of each blood pressure data point. If the abnormality threshold is set to 2.5, when the path length of a blood pressure data point is less than 2.5, it will be judged as an abnormal point and eliminated, thereby ensuring the accuracy and reliability of the data.

[0085] After the above processing, cleansed time series data is generated to ensure data accuracy and reliability. The system then performs a sliding window match on the cleansed time series data against the historical health behavior feature set. For example, a fixed-length time window is set and slid across the historical and new data, comparing the data features within the window. This match extracts data distribution shifts—the degree of difference in the distribution between the new and historical data—as well as pattern variation characteristics, such as whether new behavioral patterns have emerged or changes to existing patterns. Based on the data distribution shifts, the system adjusts the statistical distribution characteristics of the health behavior feature set, such as the mean and variance, to reflect the changes introduced by the new data. Furthermore, the pattern variation characteristics are used to expand the pattern repetition features in the activity association features, enriching the description of user behavior patterns. Finally, the updated health behavior feature set is fed into the profile generation model for training. The model adjusts its parameters based on this new feature data, further refining its understanding of the user's health status and behavioral preferences. This generates more accurate user profile text and, in turn, generates more effective health management information push strategies, thereby continuously improving the quality and accuracy of health management services.

[0086] It should be noted that when implementing the above technical solution, those skilled in the art can perform time series alignment processing based on the time series processing methods in the prior art (such as linear interpolation or cubic spline interpolation), so as to solve the problem of time stamp accuracy differences between different devices and ensure the continuity of physiological indicator data.

[0087] For the temporal convolution module of the physiological feature extraction model, we can refer to mature temporal convolutional network structures such as WaveNet or TCN to design multi-layer dilated convolutions, and combine them with existing attention mechanisms (such as multi-head attention in Transformer) to realize the correlation analysis between local trend features and global trends, thereby fully extracting periodic fluctuation features.

[0088] In the construction of behavioral association graphs, a spatial distance calculation method based on cosine similarity can be used, combined with graph clustering techniques such as the Louvain algorithm to perform community division, thereby accurately identifying user behavior patterns; for the two-way attention interaction in feature fusion, the query-key-value attention mechanism in the cross-modal alignment model can be used as a reference, and the feature aggregation process can be optimized by calculating weights through scaling dot products and introducing residual connections to ensure the deep integration of physiological and behavioral features.

[0089] The calculation of statistical distribution characteristics can dynamically update data distribution parameters based on the sliding window mean-variance method, and combine time series decomposition technology (such as STL decomposition) to separate the periodic component and the trend component, thereby enhancing the completeness of physiological state characteristics.

[0090] When updating user profiles, incremental learning methods can be used to dynamically adjust model parameters. The isolation forest algorithm can be combined to detect new data anomalies, and the dynamic time warping (DTW) algorithm can be used to calculate pattern variation characteristics, thereby continuously optimizing push strategies. Furthermore, the Z-score normalization method should be introduced during data preprocessing to unify dimensions and avoid feature extraction bias caused by unit differences in indicators such as blood oxygen and heart rate.

[0091] It can be seen that the embodiment of the present invention can generate accurate user portrait text by obtaining a health management data set including physiological indicator monitoring data and behavioral activity data, and performing effective feature extraction and fusion, and then formulate a highly targeted health management information push strategy, thereby improving the existing technology in terms of accuracy, lack of comprehensiveness and waste of network resources in health management information push.

[0092] It should also be noted that the embodiments of the present invention have constructed a complete technical solution from multi-source heterogeneous data collection, time series alignment processing, behavioral pattern mining to personalized information push by integrating specific technical means such as physiological indicator monitoring of wearable devices, behavioral semantic analysis of natural language processing, feature extraction of temporal convolutional networks and attention mechanisms, feature fusion of cross-modal alignment, push strategy generation with dynamic priority sorting, and model parameter optimization based on user feedback.

[0093] Among them, the time series alignment uses a cubic spline interpolation algorithm to solve the problem of time stamp accuracy differences among multiple devices, uses a dilated causal convolution kernel to extract local periodic trends of physiological indicators, uses a graph clustering algorithm to identify social association patterns of dietary behavior, and relies on interface services to trigger scheduled task instructions to achieve targeted information transmission of terminal devices, forming a collaborative technical system of "data perception-feature modeling-strategy generation-physical device response", which can significantly improve the push accuracy of health management information and reduce network redundant transmission. Its technical effects are reflected in optimizing the signal-to-noise ratio of physiological indicators through an outlier detection algorithm, dynamically updating user behavior features based on sliding window matching, and using incremental learning to adjust the parameters of the portrait model to adapt to real-time data distribution offsets. These technical features are deeply coupled with sensor data acquisition, computing resource allocation, and communication protocol control. They are not simply abstract levels of algorithmic rules or thinking steps. They are technical solutions that use technical means to solve information processing efficiency and resource optimization problems in specific fields. Furthermore, the embodiments of the present invention provide targeted push notifications for different users, rather than differentiated division and treatment of user groups.

[0094] The embodiment of the present invention first obtains a health management data set containing multiple health behavior records, and accurately extracts a health behavior feature set based on feature extraction processing, which can clearly present the user's health behavior characteristics; based on the feature fusion processing of the portrait generation model, it can integrate features of different dimensions to generate accurate and personalized user portrait text; the health management information push strategy generated based on the portrait text fully considers the different health conditions and behavior patterns of users, making the push content more targeted; based on this, the health management information push strategy is sent to the health management service platform for targeted push, which can ensure that users receive health management information that truly meets their needs, improve the accuracy and effectiveness of information push, and reduce the waste of network resources caused by invalid push.

[0095] See also Figure 2 As shown in FIG. 1 , the figure is a schematic diagram of the basic structure of a user portrait text generation system 200 provided by an embodiment of the present invention. The user portrait text generation system 200 includes: Processor 201; a storage device 202 having a computer program 2020 stored thereon; When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the user portrait text generation methods based on health management information push.

[0096] Based on the above, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0097] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

Claims

1. A method for generating user portrait text based on health management information push, characterized in that: include: In response to the push analysis authorization authentication information fed back by the target user, obtaining a health management data set of the target user, the health management data set including a plurality of health behavior records, each health behavior record consisting of at least one physiological indicator monitoring data and corresponding behavioral activity data; Performing feature extraction processing on the health management data set to obtain a health behavior feature set of the health behavior record, wherein the health behavior feature set includes physiological state features and activity association features; Based on a preset portrait generation model, feature fusion processing is performed on the physiological state features and the activity-related features to generate a user portrait text; A health management information push strategy is generated based on the user portrait text, and the health management information push strategy is sent to the health management service platform for targeted push operations.

2. The method according to claim 1, characterized in that The performing feature extraction processing on the health management data set to obtain the health behavior feature set of the health behavior record includes: Performing time alignment processing on the physiological indicator monitoring data in the health behavior record to generate a time-aligned physiological indicator sequence; Calling a preset physiological feature extraction model to perform periodic fluctuation analysis on the time-series aligned physiological indicator sequence to obtain physiological state features of the health behavior record, wherein the physiological state features include trend change features and abnormal fluctuation features of the physiological indicators; Performing semantic parsing on the behavioral activity data in the health behavior record to generate a structured activity tag set; A preset behavior feature extraction model is called to perform context association analysis on the structured activity tag set to obtain activity association features of the health behavior record, wherein the activity association features include spatiotemporal distribution features and pattern repetition features of the behavior activity.

3. The method according to claim 2, characterized in that The calling of a preset physiological feature extraction model to perform periodic fluctuation analysis on the time-aligned physiological indicator sequence to obtain the physiological state features of the health behavior record includes: Inputting the time-series aligned physiological indicator sequence into the temporal convolution module of the physiological feature extraction model, and extracting the local trend features of the physiological indicator sequence through multi-layer convolution kernels; Inputting the local trend features into the attention allocation module of the physiological feature extraction model, and generating a temporal attention weight based on the correlation between the local trend features of different time windows; Performing weighted aggregation on the local trend features according to the time attention weight to obtain a global trend feature; The global trend feature is spliced ​​with the statistical distribution feature of the time-series aligned physiological indicator sequence to generate the physiological state feature, the statistical distribution feature includes mean, variance and skewness coefficient, and the global trend feature includes a periodic trend component and a seasonal fluctuation component.

4. The method according to claim 3, characterized in that The calling of a preset behavior feature extraction model to perform context association analysis on the structured activity tag set to obtain activity association features of the health behavior record includes: Mapping the structured activity tag set to a pre-generated behavior semantic space to generate an initial behavior semantic vector; Invoking the spatial association module of the behavior feature extraction model to calculate the spatial distance features between the initial behavior semantic vectors, and generating a behavior association graph based on the spatial distance features; Performing community division processing on the behavior association graph to obtain multiple behavior community subgraphs, each behavior community subgraph corresponding to a behavior pattern category; The topological structure features and node distribution features of each behavioral community subgraph are extracted, and the topological structure features and the node distribution features are fused to generate the activity association features. The topological structure features include community density and centrality indicators, and the node distribution features include label distribution uniformity and time overlap.

5. The method according to claim 1, characterized in that The preset portrait generation model is based on which the physiological state features and the activity-related features are subjected to feature fusion processing to generate a user portrait text, including: Inputting the physiological state features and the activity-related features into the cross-modal alignment module of the portrait generation model to generate a feature alignment matrix; performing bidirectional attention interaction processing on the physiological state feature and the activity-related feature based on the feature alignment matrix to generate a fused feature vector; Input the fused feature vector into the text generation module of the portrait generation model, and expand the decoder layer by layer to generate semantic paragraphs of the user portrait text, wherein the semantic paragraphs include a health status summary, a behavior preference analysis, and a summary of intervention recommendations; The semantic paragraphs are subjected to coherence verification processing, redundant descriptions are deleted, and the order of the paragraphs is adjusted to generate the user portrait text.

6. The method according to claim 5, characterized in that The performing bidirectional attention interaction processing on the physiological state feature and the activity-related feature based on the feature alignment matrix to generate a fused feature vector includes: Calculating a first attention weight by using the physiological state feature as a query vector and the activity-related feature as a key vector and a value vector; performing weighted aggregation on the activity-related features according to the first attention weight to obtain a first aggregated feature; Calculating a second attention weight by using the activity-related feature as a query vector and the physiological state feature as a key vector and a value vector; Performing weighted aggregation on the physiological state features according to the second attention weight to obtain a second aggregated feature; The first aggregated feature and the second aggregated feature are concatenated to generate the fused feature vector.

7. The method according to claim 1, characterized in that The user portrait text is used to describe the health management demand characteristics and potential behavior preference characteristics of the target user. The health management information push strategy generated according to the user portrait text includes: Perform keyword extraction processing on the user portrait text according to the health management demand characteristics in the user portrait text to generate a health demand tag set; Matching and searching the health demand tag set with a pre-generated health knowledge base to obtain a candidate health management information set; Sorting the candidate health management information set according to the potential behavioral preference characteristics in the user portrait text to generate a priority queue; A preset number of target health management information is selected based on the priority queue, and a push time strategy and a push channel strategy are configured to generate the health management information push strategy.

8. The method according to claim 7, characterized in that The step of sorting the candidate health management information set according to the potential behavior preference features in the user portrait text to generate a priority queue includes: Extracting time period features describing behavior time preferences and device type features describing channel usage preferences from the user portrait text; Calculating a matching weight between the content topic of each piece of information in the candidate health management information set and the time period feature; Calculating the compatibility weight between the push channel adaptability of each piece of information in the candidate health management information set and the device type characteristics; The candidate health management information sets are arranged in descending order according to a weighted sum of the matching weight and the conformity weight to generate the priority queue.

9. The method according to claim 1, characterized in that The sending of the health management information push policy to the health management service platform for targeted push operation includes: Analyze the target health management information, push time strategy and push channel strategy in the health management information push strategy; Generate a scheduled task instruction according to the push time strategy, and cache the target health management information in a push queue of the health management service platform; When the execution time of the scheduled task instruction is reached, the interface service of the health management service platform is called, and the target health management information is sent to the terminal device of the target user according to the push channel strategy; Monitor user interaction data of the target health management information, and update potential behavior preference features in the user portrait text according to the user interaction data.

10. A user portrait text generation system, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the user portrait text generation method based on health management information push as described in any one of claims 1-9.

Images