Information recommendation method, device, apparatus, storage medium, and program product

Abstract

The embodiment of the present specification provides an information recommendation method, device, apparatus, storage medium and program product. In the embodiment of the present application, by maintaining medical health portrait data including initial theme and copy style information, using a theme generation large model to perform semantic expansion on the initial theme, a plurality of target themes are dynamically constructed; a copy generation large model is called, and corresponding candidate medical health recommendation copy is generated for each target theme based on the copy style information; a target medical health recommendation copy is selected from among them for pushing, so that the information recommendation is converted from the unified distribution of the static content library to the personalized customization of thousands of people with thousands of faces, the accuracy of the recommended content is significantly improved, and the differentiated health management needs are met.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to an information recommendation method, device, apparatus, storage medium, and program product. Background Technology

[0002] In the healthcare field, the integrated application of Artificial Intelligence (AI) and Large Models (LM) is driving the development of services such as intelligent consultation, report interpretation, disease management, and health information delivery. Among these, the information delivery service aims to provide health knowledge, risk warnings, and personalized suggestions based on user needs.

[0003] Currently, the traditional approach is to pre-build a health content library and push relevant content to users through a matching mechanism. However, the content recommended in this way is not only prone to being monotonous and repetitive, but may also lack accuracy and fail to meet the personalized health management needs of each individual. Summary of the Invention

[0004] This specification provides an information recommendation method, device, apparatus, storage medium, and program product to provide personalized recommendation services to a variety of users, meeting the "personalized" health management needs of each individual.

[0005] This specification provides an information recommendation method, which involves: constructing a medical and health profile of the target user based on multi-source behavioral data and health record data of the target user in medical and health applications; the medical and health profile data including at least initial topics of interest to the target user and copywriting style information; using a topic generation model to semantically expand the initial topics based on the medical and health profile data to obtain multiple target topics; using a copywriting generation model to generate candidate medical and health recommendation texts corresponding to each of the multiple target topics according to the copywriting style information; selecting a target medical and health recommendation text from the candidate medical and health recommendation texts and recommending the target medical and health recommendation text to the target user.

[0006] This specification also provides an information recommendation device, comprising: a construction module, a topic generation module, a copywriting generation module, and a recommendation module; the construction module is used to construct medical and health profile data of the target user based on multi-source behavioral data and health record data of the target user in medical and health applications, the medical and health profile data including at least initial topics of interest to the target user and copywriting style information; the topic generation module is used to use a topic generation model to semantically expand the initial topics based on the medical and health profile data to obtain multiple target topics; the copywriting generation module is used to use a copywriting generation model to generate candidate medical and health recommendation copywriting corresponding to each of the multiple target topics according to the copywriting style information; the recommendation module is used to select target medical and health recommendation copywriting from the candidate medical and health recommendation copywriting and recommend the target medical and health recommendation copywriting to the target user.

[0007] This specification also provides an electronic device, including: a memory and a processor; the memory is used to store one or more computer instructions; the processor is used to execute one or more computer instructions to perform the steps in the method provided in this specification.

[0008] This specification also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the method provided in this specification.

[0009] This specification also provides a computer program product, including: a computer program / instructions, which, when executed by a processor, can implement the steps of the method provided in this specification.

[0010] In this embodiment, by maintaining medical and health profile data including initial theme and copywriting style information, a large theme generation model is used to expand the initial theme and dynamically construct multiple target themes; a large copywriting generation model is called to generate corresponding candidate medical and health recommendation copywriting for each target theme based on copywriting style information; and target medical and health recommendation copywriting is selected and pushed, so that information recommendation is transformed from uniform distribution of static content library to personalized customization of each person, which significantly improves the accuracy of recommended content and meets the needs of differentiated health management. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating an exemplary embodiment of the information recommendation method provided in this specification.

[0012] Figure 2 This is a flowchart illustrating another information recommendation method provided in one embodiment of this specification.

[0013] Figure 3 This is a schematic diagram illustrating the internal execution of an information recommendation stage according to an embodiment of this specification.

[0014] Figure 4 This is a schematic diagram of an information recommendation device provided for an exemplary embodiment of this specification.

[0015] Figure 5 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this specification. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0017] It should be noted that, in the cases involving user information in this application's embodiments, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. The various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0018] Traditional information push solutions pre-build a health content library and push relevant content to users through a matching mechanism. This approach often results in content that is monotonous, highly repetitive, and lacks precision, failing to meet the personalized health management needs of each user. In the embodiments described in this specification, by maintaining medical and health profile data including initial topics and text style information, a large-scale topic generation model is used to semantically expand the initial topics, dynamically constructing multiple target topics. A large-scale text generation model is then invoked to generate corresponding candidate medical and health recommendation texts for each target topic based on text style information. The target medical and health recommendation text is then selected for push, transforming information recommendation from a static, uniform distribution based on a content library to personalized customization for each user. This solves the problem of highly repetitive content recommendations for different users, significantly improves the accuracy of recommended content, and meets differentiated health management needs.

[0019] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating an exemplary embodiment of an information recommendation method provided in this specification, such as... Figure 1 The method shown includes: 101. Based on the target user's multi-source behavioral data and health record data in medical and health applications, construct the target user's medical and health profile data. The medical and health profile data shall include at least the initial topics and copywriting style information that the target user is interested in.

[0021] 102. Utilize topic-based large-scale model generation. Based on medical and health profile data, semantic expansion is performed on the initial topic to obtain multiple target topics.

[0022] 103. Utilize a copywriting generation model to generate candidate medical and health recommendation copy for each of the target themes based on copywriting style information.

[0023] 104. Select the target medical and health recommendation copy from the candidate medical and health recommendation copy and recommend the target medical and health recommendation copy to the target users.

[0024] In the embodiments described in this specification, the healthcare application is an artificial intelligence (AI)-based healthcare application, such as a native application developed based on AI. This application is an intelligent health assistant with capabilities including natural language understanding, multimodal data processing, medical logic reasoning, and personalized generation. It is a software system that can assist different users in disease prevention, health monitoring, diagnostic assistance, treatment management, rehabilitation guidance, and the development of healthy lifestyles. The following description uses information recommendation for target users within the healthcare application as an example.

[0025] In the embodiments of this specification, multi-source behavioral data includes: dialogue behavior data and operational behavior data, etc. Dialogue behavior data includes, but is not limited to: historical multi-turn dialogue records between the target user and the healthcare application, etc. Operational behavior data includes, but is not limited to: click behavior on notification information, browsing time for content, and frequently used functional modules in the healthcare application, etc. Among them, health record data includes, but is not limited to: the target user's basic information (e.g., nickname, geographical location), medical history, physical examination information, examination results, and medication record information, etc.

[0026] It should be noted that the multi-source behavioral data involved in the embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of the relevant data comply with relevant laws, regulations, and standards. Among them, the above... Figure 1 During the execution of the method shown, all the above data are de-identified and / or encrypted, and the processing flow is not visible to the outside world, ensuring the security of data throughout the entire chain.

[0027] Multi-source behavioral data is primarily used to capture and quantify the health management preferences and immediate needs of target users in real time. Dialogue behavior data reflects the specific health issues or information needs currently being addressed by the target user; operational behavior data (such as clicks, browsing duration, and frequently used functional modules) reveals the target user's acceptance level, interest intensity, and usage habits regarding different health content. These data, working together, can identify fine-grained interest tendencies from the target user's dynamic interactions, providing real-time evidence for subsequently deleting target topics and ensuring that the constructed target topics align with the target user's current actual concerns.

[0028] Among these, health record data is primarily used to build the foundation for a user's long-term, static health profile. By integrating basic information, medical history, physical examination and medication records, a comprehensive understanding of the target user's health status, risk factors, and health management stage can be obtained. This underlying data provides logical support for setting the "initial theme and copywriting style information," ensuring that the subsequently generated medical and health recommendation copy not only conforms to the target user's objective health reality (such as chronic disease management needs) but also adheres to medical rigor.

[0029] In summary, these two types of data, from the perspectives of dynamic behavioral preferences and static health essence respectively, lay a solid data foundation for generating medical and health profiles of target users and for the subsequent transformation from "uniform distribution" to "personalized recommendations," thereby significantly improving the matching degree between recommended content and individual health needs.

[0030] In the embodiments of this specification, based on multi-source behavioral data and health record data of the target user in medical and health applications, medical and health profile data of the target user is constructed, transforming fragmented raw data into structured knowledge that can be understood and reasoned about by machines. The medical and health profile data includes at least: initial topics of interest to the target user and copywriting style information. Initial topics can be health topics raised by the user based on their own perceptions, interests, or needs. Initial topics may include, but are not limited to: topics consulted by the target user in historical periods, preferably topics consulted recently (e.g., in the last week or month). For example, initial topics may include, but are not limited to: cardiovascular and cerebrovascular health, sleep disorders, etc. Copywriting style information is used to reflect the target user's style preferences. For example, from the perspective of tone and intonation, copywriting style information includes, but is not limited to: professional and rigorous, warm and caring, warning and serious, and lighthearted and humorous, etc.; from the perspective of audience adaptation, copywriting style information includes, but is not limited to: simple style, easy to understand, and in-depth explanation of standard terminology, etc. From the perspective of content structure, copywriting style information includes, but is not limited to: narrative stories, question-and-answer dialogues, checklists, data reports, and step-by-step instructions, etc.

[0031] The implementation method for constructing the target user's medical and health profile data based on multi-source behavioral data and health record data in medical and health applications is not limited. For example, regarding the extraction of the initial topic, the health record data is structured and parsed to map the user's medical history, abnormal physical examination indicators, and medication records into static health risk tags. For example, "impaired fasting glucose" and "family history of diabetes" can be directly transformed into the initial topic of "prediabetes intervention." Simultaneously, dynamic corrections are made using multi-source behavioral data. By analyzing the user's recent search keywords, frequently viewed functional modules (such as frequent viewing of "insomnia treatment"), click behavior on specific notifications, and pain points actively inquired about in historical conversations, the target user's current immediate health needs are captured. The aforementioned static risk tags are then integrated and deduplicated with dynamic interest hotspots to ultimately determine an initial topic that conforms to medical and health logic and aligns with the target user's current concerns. This initial topic serves as a seed for the larger model to extend vertically, derive horizontally, and expand preventatively. It is no longer limited to a static content library and can continuously generate personalized content that is timely and relevant to the current scenario.

[0032] For example, extracting copywriting style information primarily relies on in-depth analysis of dialogue behavior data and operational habits. Natural language processing is used to analyze the target user's historical multi-turn dialogue records with healthcare applications, analyzing their vocabulary habits (preference for professional medical terminology or colloquial language), sentence length (preferring concise instructions or detailed explanations), and emotional inclination (leaning towards rational objectivity or emotional encouragement). Furthermore, by combining the target user's operational behavior within the healthcare application, such as the average time spent reading long articles and whether they prefer watching videos over text, the target user's preferences for content format are further refined. These characteristics are quantified into specific copywriting style information, such as "professional and rigorous," "friendly and caring," or "concise and to the point," and this copywriting style information is injected as a key parameter into the healthcare profile data. By maintaining copywriting style information, it is ensured that the generated content is highly aligned with the target user's cognitive level and emotional preferences. For professional users, using rigorous terminology establishes a sense of authority; for elderly or anxious users, using friendly and colloquial language lowers the comprehension threshold and alleviates anxiety. This personalized approach makes target users feel truly understood, making them more willing to read and follow health advice, thus significantly improving the actual effectiveness of health management.

[0033] The generated medical and health profile data is no longer a collection of scattered database fields, but a structured object containing a clear "initial topic" and specific "copywriting style" constraints. This object directly serves as the input context for subsequent calls to the topic generation model and copywriting generation model, ensuring that the generated recommendations are medically accurate and safe, and personalized in terms of expression.

[0034] In addition to the above, medical and health profile data can also include attribute description information of the target user. For example, attribute description information can be directly obtained from health record data and integrated into a structured form. Attribute description information may include, but is not limited to: natural attributes, social attributes, life stage, medical history, health status, psychological state, lifestyle and behavioral data, etc. Natural attributes may include, but are not limited to: education level, height, and weight, etc. Social attributes may include, but are not limited to: place of residence, family structure, marital status, and reproductive history, etc. Figure 3The illustration uses initial topic and copywriting style information as an example of medical and health profile data, but it is not limited to this. Attribute description information (such as medical history, medication records, allergy history, etc.) provides strict medical and health constraints for the large model. When generating content, these attributes can automatically filter out suggestions harmful to the target user (e.g., avoiding recommending a high-protein diet to a user with renal insufficiency), ensuring that all generated target topics and copywriting are based on the target user's actual physiological condition. This not only avoids the medical risks that general recommendations may bring, but also allows push priorities to be intelligently sorted according to the urgency of medical and health issues, focusing limited attention resources on key health interventions.

[0035] In the embodiments of this specification, to ensure the accuracy of the content pushed during the information recommendation process to target users, a large topic generation model is maintained. Using this model, based on medical and health profile data, the initial topics are semantically expanded, thereby dynamically generating multiple target topics with high adaptability. This significantly enriches the matching dimensions of the recommended content, provides more target topics suitable for the target users, and enables the prediction and guidance of user needs. Figure 2 and Figure 3 As shown. In Figure 2 In this context, the above process belongs to the information recommendation stage.

[0036] The topic generation large-scale model is a specialized vertical model designed to generate target topics. It's a deep learning model trained on a massive corpus with a huge number of parameters. Given a text (i.e., context), the large model attempts to predict the most likely next word. For example, the topic generation large-scale model could be a large language model. The number of parameters in the large model exceeds a set threshold, which can be in the millions, billions, hundreds of billions, or even trillions, depending on the application requirements. It's not limited to the range of model parameters; for example, depending on the needs, the number of parameters can be below a million. During training, these parameters are continuously adjusted and optimized based on the difference between the large model's predictions and the labeled results to improve its prediction accuracy. In some embodiments, the large model typically employs an advanced neural network architecture, such as a Transformer architecture, to build its structure. This architecture enables the large model to capture complex patterns in the text and handle long-distance dependencies. After sufficient training, the large model possesses powerful generative capabilities, producing coherent and context-appropriate text content based on given prompts.

[0037] In this application embodiment, the implementation method of semantic expansion of the initial topic is not limited. For example, semantic expansion of the initial topic can be performed on a single dimension or on multiple dimensions. Taking semantic expansion of the initial topic on a single dimension as an example, it can be performed on the semantic depth, referred to as vertical extension of semantic depth; or it can be performed on the semantic breadth, referred to as horizontal derivation of semantic breadth; or it can be performed on the temporal dimension for preventive expansion, referred to as preventive expansion of the temporal dimension. Taking semantic expansion of the initial topic on multiple dimensions as an example, it can be performed on the initial topic with vertical extension of semantic depth and horizontal derivation of semantic breadth; or, for example, with horizontal derivation of semantic breadth and preventive expansion of the temporal dimension; or, for example, with vertical extension of semantic depth, horizontal derivation of semantic breadth, and preventive expansion of the temporal dimension respectively, to obtain multiple target topics.

[0038] Vertical extension involves starting from the surface-level topics or initial understanding raised by the target users and delving deeper into causal chains, pathological mechanisms, and diagnostic and treatment levels, focusing on "why" and "how." This can solidify the target users' understanding and provide them with advanced guidance. For example, if the initial topic is daily dietary recommendations for diabetes, vertical extension could yield target topics including, but not limited to: the molecular mechanisms of insulin resistance, dosage adjustment strategies for diabetes medications, and the timing of treatment for diabetic retinopathy.

[0039] Horizontal derivation involves expanding upon an initial topic to related fields, life scenarios, or multidisciplinary perspectives. It focuses on identifying "what other topics are possible" and "what related topics are relevant," providing a multi-perspective knowledge network. For example, if the initial topic is daily dietary recommendations for diabetes, horizontally derived target topics could include, but are not limited to: psychological intervention for diabetes complicated by depression, common dietary planning and cooking techniques for diabetes, and care guidelines for diabetes and periodontal health.

[0040] The preventative extension focuses on shifting the timeline forward, expanding from the initial topic to risk prediction, early warning, and full life-cycle management. It emphasizes "what will the future hold" and "how to prevent problems before they occur," utilizing historical trends in healthcare profile data to predict future risks. While the initial topic is daily dietary recommendations for diabetes, the preventative extension can lead to target topics including, but not limited to: cardiovascular risk assessment for the next ten years, family history-based prediabetes reversal programs, and full life-cycle strategies from childhood obesity intervention to adult diabetes prevention.

[0041] In the embodiments of this specification, to ensure that not only the recommended content is adapted to the target users, but also that the pushed information's expression style matches the users' aesthetic and cognitive habits, this embodiment also maintains a large-scale copywriting generation model. Using this model, based on copywriting style information, multiple candidate medical and health recommendation texts corresponding to each target topic are generated. This transforms the target topic into candidate medical and health recommendation texts adapted to the user in terms of writing style, structure, and expression. For example, one target topic may correspond to one or more candidate medical and health recommendation texts. Figure 2 and Figure 3 As shown. In Figure 2 In this context, the above process belongs to the information recommendation stage.

[0042] The copywriting generation big model is a deep learning model with massive parameters trained on a large-scale corpus. Given a text (i.e., context), the big model attempts to predict the most likely next word. For example, the copywriting generation big model can be implemented as a large language model. For an introduction to the big model, please refer to the aforementioned examples, which will not be repeated here.

[0043] In the embodiments of this specification, a target medical and health recommendation is selected from candidate medical and health recommendation texts and recommended to the target user. The timing of this selection and recommendation is not limited. For example, a target push time point can be selected from any time period, and the target medical and health recommendation text can be selected and recommended to the target user. Alternatively, when the target push time period arrives, a target push time point can be selected, for example, a time point can be randomly selected from the target push time period, and the target medical and health recommendation text can be selected and recommended to the target user. The target push time period can be a preset time period, including but not limited to: 7:00-8:00, 8:00-9:00, 9:00-10:00, 11:00-12:00, 17:00-18:00, 18:00-19:00, 20:00-21:00, and 21:00-22:00. exist Figure 3 The document demonstrates a timed triggering process, specifically referring to recommending target medical and health information to target users.

[0044] The implementation method for selecting the target medical and health recommendation from candidate medical and health recommendations is not limited. In one example, the target medical and health recommendation can be selected based on the target push time period. For example, if the target push time period is from 7:00 to 8:00 AM, candidate medical and health recommendations in the categories of "breakfast suggestions" or "morning exercise" are prioritized as the target medical and health recommendation; if the target push time period is before bedtime, such as 10:00 PM or 11:00 PM, candidate medical and health recommendations in the categories of "sleep improvement" or "tomorrow's medication reminder" are selected as the target medical and health recommendation. In another example, if the target push time period is before the target medical and health recommendation period, candidate medical and health recommendations that match the topic of the target user's last consultation are selected as the target medical and health recommendation. For example, if the target user consulted on a topic related to "headache" in the medical and health application just before the push, candidate medical and health recommendations related to "headache relief" can be selected as the target medical and health recommendation.

[0045] The method for recommending the target medical and health information to the target user is not limited. For example, a notification message could pop up on the lock screen or notification bar of the device containing the medical and health application; this notification message could be a summary of the target medical and health recommendation. Another example is generating a new message in the "Messages" or "Notifications" section within the medical and health application, containing the target medical and health recommendation; the target user can see this recommendation when they open the application. Yet another example is displaying the target medical and health recommendation as a prominent card at the top of the homepage or in the "Daily Health Advice" section when the target user opens the application. Still another example is reading the target medical and health recommendation aloud via voice.

[0046] In the embodiments of this specification, medical and health profile data including initial topic and copywriting style information is maintained. The initial topic is semantically expanded using a topic generation model to dynamically construct multiple target topics. The copywriting generation model is called to generate corresponding candidate medical and health recommendation copy for each target topic based on the copywriting style information. The target medical and health recommendation copy is selected and pushed, so that information recommendation is transformed from uniform distribution of static content library to personalized customization of each person, which significantly improves the accuracy of recommended content and meets the needs of differentiated health management.

[0047] In one optional embodiment, the implementation method for selecting the target medical and health recommendation text from candidate medical and health recommendation texts is not limited. In one example, to achieve non-repetitive and highly relevant information push, an independent buffer pool is maintained for different users to store candidate medical and health recommendation texts generated for that user. For ease of distinction and description, the buffer pool corresponding to the target user is called the target buffer pool. One implementation method for selecting the target medical and health recommendation text from candidate medical and health recommendation texts when the target push period arrives includes: adding candidate medical and health recommendation texts corresponding to multiple target topics to the target buffer pool corresponding to the target user as schedulable resources; when the target push period arrives, selecting the target medical and health recommendation text from the target buffer pool and recommending the target medical and health recommendation text to the target user. If the push is successful, the selected target medical and health recommendation text is deleted from the target buffer pool, which ensures that the pushed information is non-repetitive and highly relevant, and also reduces the cost and latency problems caused by high-frequency calls to large models. In addition, secondary filtering is achieved through the target buffer pool to ensure that the pushed content is adapted to the target push time point, thereby improving the user's click-through rate. Figure 2 and Figure 3 The example shown here is for the target buffer pool.

[0048] This can be achieved by directly adding candidate medical and health recommendation texts to the target buffer pool corresponding to the target user, or by adding annotations to each candidate medical and health recommendation text, such as the reason, theme type, and writing style information. For example, for a user who prefers a concise list style and has a history of hypertension, a theme titled "Five Dos and Three Don'ts for Hypertension Diet in Summer" could be generated, with the reason labeled as "Seasonal Deepening," the theme type as "Deepen," and the writing style information as "List Style." Or, for a white-collar user who likes visual content, a theme titled "5-Minute Neck Rescue Guide for the Office" could be generated. The reason given was "occupational risk prevention," the theme type was "diverge," and the copywriting style was "simple with emojis."

[0049] In one optional embodiment, the implementation method of using a large-scale topic generation model to semantically expand an initial topic based on medical and health profile data to obtain multiple target topics is not limited. In one example, the large-scale topic generation model corresponds to a medical and health knowledge graph, which includes multiple topic nodes. These topic nodes have vertical hierarchical relationships, horizontal hierarchical relationships, and temporal relationships. Vertical hierarchical relationships represent parent-child dependencies between topic nodes, reflecting hierarchical relationships. Horizontal hierarchical relationships represent parallel, related, or comorbid relationships; for example, the relationship between two topic nodes belonging to the same parent node. Temporal relationships represent causal or evolutionary relationships along a timeline. Each topic node maintains a topic tag and a content summary. The topic tag is the identifier or core concept name of the topic node, used to uniquely identify the medical and health entity or concept represented by the node. The content summary is a structured or semi-structured textual description of the medical knowledge contained in the topic node.

[0050] The aforementioned vertical hierarchical relationship corresponds to the vertical extension in semantic depth, the horizontal hierarchical relationship corresponds to the horizontal derivation in semantic breadth, and the temporal association relationship corresponds to the preventive expansion in the time dimension. Based on this, a method using a large-scale topic generation model to semantically expand an initial topic based on medical and health profile data to obtain multiple target topics is as follows: Using the large-scale topic generation model, based on the medical and health profile data, the initial topic is subjected to vertical extension in semantic depth, horizontal derivation in semantic breadth, and preventive expansion in the time dimension to obtain multiple target topics. Specifically, the implementation involves: semantically parsing the medical and health profile data to obtain the parsing result, which includes the initial topic; using the initial topic as the query benchmark, traversing nodes in the medical and health knowledge graph to locate the first topic node corresponding to the initial topic; using the first topic node as the anchor point, querying along the vertical hierarchical relationship, horizontal hierarchical relationship, and temporal association relationship of the first topic node in the medical and health knowledge graph to obtain multiple related topic nodes; and calling the large-scale topic generation model to generate multiple target topics based on the topic tags and content summaries of each of the multiple related topic nodes.

[0051] The process involves using an initial topic as the query baseline, traversing nodes within the medical and health knowledge graph to locate the first topic node corresponding to the initial topic. This maps the initial topic to the structured medical and health knowledge graph, determining the starting point for the graph traversal. Subsequently, using this first topic node as an anchor point, expansion occurs along three dimensions or paths. Using the first topic node as an anchor point, preventative topic expansion is performed within the medical and health knowledge graph along the vertical, horizontal, and temporal hierarchical relationships of the first topic node, resulting in multiple related topic nodes along these three paths.

[0052] The implementation method for generating multiple target topics based on the topic tags and content summaries of multiple related topic nodes using the large-scale topic generation model is not limited. For example, the topic tags and content summaries of multiple related topic nodes can be input into the large-scale topic generation model. Semantic fusion and intent reconstruction can be performed on the topic tags and content summaries of each related topic node to generate a target topic. That is, the structured knowledge information (topic tags and content summaries) is transformed into a user-oriented, attractive, and action-guiding natural language topic title. Alternatively, a target topic can be generated based on the topic tags and content summaries of at least two related topics on each path, without limitation.

[0053] Optionally, using the first topic node as an anchor point, queries can be performed in the medical and health knowledge graph along the vertical, horizontal, and temporal hierarchical relationships of the first topic node to obtain multiple related topic nodes. The implementation method is not limited. In one example, the parsing results of the medical and health profile data also include: topic context information corresponding to the initial topic, and steady-state feature information and environmental context information corresponding to the target user. This information can provide a reasoning basis for the topic generation model, enabling deep matching within the large model.

[0054] Thematic context information refers to semantically relevant and logically connected supplementary medical and health background information surrounding the initial topic. It is an explanatory, supporting, or extended description of the initial topic in the current user scenario. For example, if the initial topic is "high fasting blood glucose", its thematic context information may include, but is not limited to: {clinical significance: suggesting possible insulin resistance or prediabetes}, {related indicators: glycated hemoglobin, 2-hour postprandial blood glucose}, {typical symptoms: thirst, polyuria, fatigue (even if the user does not actively mention it)}; {common causes: high-carbohydrate diet, lack of exercise, insufficient sleep}; {knowledge source: Chinese Guidelines for the Prevention and Treatment of Type 2 Diabetes, Edition X}.

[0055] Homeostatic characteristics refer to the relatively stable attributes of a target user over a longer period of time, which are not easily changed in the short term. This type of information reflects the target user's basic health status and is characterized by low-frequency updates and high predictive value. For example, homeostatic characteristics may include, but are not limited to: genetic history, history of chronic diseases, physical fitness indicators, type of work, and lifestyle habits.

[0056] Among them, environmental context information refers to the external dynamic situational factors in which the target user is situated, including but not limited to: life stage, lifestyle, social environment, recent events, epidemics, seasonal changes, and other factors that have a significant impact on health.

[0057] Based on this, a method is proposed that uses the first topic node as an anchor point to query along the vertical, horizontal, and temporal hierarchical relationships of the first topic node in the medical and health knowledge graph to obtain multiple related topic nodes, including: Based on the topic context information of the initial topic, in the medical and health knowledge graph, the child nodes corresponding to the first topic node are queried along the vertical hierarchical relationship as the vertically extended related topic nodes; based on the steady-state feature information, in the medical and health knowledge graph, sibling nodes that share the same parent node with the first topic node are queried along the horizontal hierarchical relationship as the horizontally derived related topic nodes; based on the environmental context information, in the medical and health knowledge graph, nodes that have a causal relationship with the first topic node are queried along the temporal relationship as the preventive extended related topic nodes.

[0058] For example, in a vertical extension path, the child nodes corresponding to the first topic node are obtained, the similarity between the topic context information and each child node is calculated, and child nodes with a similarity higher than a set similarity threshold (e.g., 80%) are selected as related topic nodes for vertical extension. As another example, in a horizontal extension path, sibling nodes sharing the same parent node as the first topic node are obtained, the similarity between the steady-state feature information and each sibling node is calculated, and sibling nodes with a similarity higher than a set similarity threshold (e.g., 91%) are selected as related topic nodes for horizontal extension. Yet another example, in a preventative extension path, other topic nodes with causal relationships to the first topic node are obtained, the similarity between the environmental context information and other topic nodes is calculated, and other topic nodes with a similarity higher than a set similarity threshold (e.g., 90%) are selected as related topic nodes for preventative extension.

[0059] In one optional embodiment, the implementation method of using a copywriting generation model to generate candidate medical and health recommendation copy for multiple target topics based on copywriting style information is not limited. In one example, medical and health profile data and multiple target topics are input into the copywriting generation model. For any target topic, user context information associated with the target topic is parsed from the medical and health profile data. Semantic expansion and content generation are performed on the target topic using copywriting style information as constraints to obtain push copy. Based on the push copy and user context information, medical and health query information is generated. The medical and health query information is input into a medical and health question-and-answer model associated with the target application for answer generation. The push copy and the medical and health query information are used as candidate medical and health recommendation copy. Figure 2 The following diagram illustrates targeted healthcare recommendation copy, including push notifications and healthcare query information.

[0060] In this context, user context information associated with the target topic refers to a subset of structured or semi-structured data from the healthcare profile data focused around that topic. For example, if the target topic is exercise recommendations, the user context information extracted from the healthcare profile data may include, but is not limited to, average daily steps, sedentary time, and heart rate variability trends. As another example, if the target topic is sleep disorder improvement, the user context information extracted from the healthcare profile data may include, but is not limited to, average sleep onset time, number of nighttime awakenings, and sleep depth.

[0061] The process involves semantic expansion and content generation of the target topic, constrained by copywriting style information, to produce the push notification copy. Semantic expansion refers to deepening, refining, or contextualizing the connotation of the target topic while preserving its core semantics. This can be done by logically expanding the implicit information dimensions of the target topic based on medical and health knowledge. Content generation, building upon the "information skeleton" constructed by semantic expansion, utilizes natural language processing capabilities to transform it into fluent, complete text that conforms to the specific copywriting style. This step involves not only vocabulary selection, sentence structure, and paragraph arrangement but also the integration of emotional warmth, calls to action, or cognitive guidance.

[0062] The push notification text is provided to the target users so they can access it. To provide a seamless interactive experience and connect information push with medical dialogue, medical and health query information corresponding to the push notification text can be pre-generated. This medical and health query information serves as the first question in the medical dialogue. When the target user accesses the push notification text, this medical and health query information is provided to the medical and health question-and-answer model associated with the target application. This allows the medical and health question-and-answer model to generate answers, ensuring that it accurately understands the push notification text's intent and provides personalized medical and health responses.

[0063] Based on this, healthcare queries can be generated using push notifications and user context information. For example, push notifications and user context information can be structurally integrated to generate a healthcare query based on a comprehensive healthcare Q&A model. This query might be identified by a fixed prefix (e.g., "[Push]") and embed two key elements: health-related keywords or phrases extracted from the push notification, and user context information (e.g., chronic disease history). This design ensures that when a target user accesses the Q&A page within a healthcare application after receiving the push notification, the application can immediately invoke the comprehensive healthcare Q&A model to generate highly accurate and personalized in-depth answers.

[0064] The medical and health question-answering big model is a deep learning model with massive parameters trained on a large-scale corpus. Given a text (i.e., context), the big model attempts to predict the most likely next word. For example, the medical and health question-answering big model can be a large language model; for an introduction to large language models, please refer to the aforementioned examples, which will not be repeated here.

[0065] Optionally, the implementation method for recommending target medical and health recommendations to target users is not limited. For example, the target push notification text included in the target medical and health recommendation text may be provided to the target user. In addition, it may include: responding to user actions on the target push notification text, inputting the target medical and health query information from the target medical and health recommendation text into a large-scale medical and health question-and-answer model to generate answers, thereby obtaining medical and health answer information; and displaying the medical and health answer information on the question-and-answer page of the target application, such as... Figure 2 As shown.

[0066] The way the targeted medical and health recommendation text, including the targeted push text, is delivered to the target users varies, and the user actions in response to the targeted push text also differ. For example, if the targeted push text is delivered to the target users in the form of a text notification, the user action in response to the targeted push text could be a click on the text notification. As another example, if the targeted push text is delivered to the target users in the form of voice, the user action in response to the targeted push text could be a read-aloud action on the targeted push text.

[0067] In addition, Figure 2 The document also demonstrates the process of conducting security screening on target medical and health recommendation texts during the push notification phase. Security screening is used to identify false advertising, erroneous advice, and illegal content, ensuring the rigor and reliability of the target medical and health recommendation texts. For example, the target medical and health recommendation texts can be input into a large-scale security screening model. This large-scale security screening model can be a large language model; for details, please refer to the aforementioned example, which will not be repeated here.

[0068] The internal structure of the copywriting generation model is not limited. Below is an example: the copywriting generation model includes a parsing module, a copywriting content generation module, a query intent construction module, and a candidate copywriting assembly module.

[0069] The parsing module is used to extract user context information related to any target topic from medical and health profile data. For example, it is implemented using a sentence-level bidirectional encoder representation (Sentence-BERT) structure.

[0070] The copywriting content generation module is used to semantically expand and generate content for a target topic, constrained by copywriting style information, to produce push copy. For example, it can be implemented using a decoder-only Transformer architecture or a pure decoder Transformer architecture.

[0071] The query intent building module is used to generate medical and health query information based on push notifications and user context information. For example, it can be implemented using a sequence-to-sequence rewriting framework.

[0072] The candidate copy assembly module is used to combine push notifications and medical and health query information as candidate medical and health recommendation copy. This module has no parameterized logical assembly units, does not rely on neural networks, and performs structured data encapsulation.

[0073] In one optional embodiment, to push "appropriate information" to target users at the "appropriate time," thereby significantly improving the effectiveness of health interventions and user engagement, a push timing prediction model is used to analyze the historical activity data of target users, accurately identify the high-frequency login, reading, and interaction periods of target users, and calculate the optimal push time period for target users, i.e., the push time window.

[0074] Among them, the push timing prediction big model is a deep learning model with massive parameters trained on a large-scale corpus. Given a text (i.e., context), the big model attempts to predict the most likely next word. For example, the medical and health question-answering big model can be a big language model. For an introduction to big language models, please refer to the aforementioned examples, which will not be repeated here.

[0075] Specifically, medical and health profile data is input into a large-scale push notification timing prediction model. Semantic parsing is performed on the medical and health profile data to obtain temporal behavioral information and attribute description information. Frequency statistics and time period aggregation are performed on the temporal behavioral information to obtain candidate push notification time periods for the target user. Using attribute description information as a weighting adjustment factor, an attention mechanism is used to refine the candidate push notification time periods to obtain the target push notification time period for the target user. Figure 3 The example demonstrates how a large-scale model for predicting push timing generates target push periods.

[0076] Among them, time-series behavioral information refers to user interaction or physiological behavior records with clear timestamps extracted from medical and health profile data, reflecting the target user's activity patterns and habit rhythms at different points in time. For example, time-series behavioral information may include, but is not limited to: application usage behavior, wearable device data, health intervention response records, and life event logs. Application usage behavior may include, but is not limited to: the time when the health app is opened daily, the time period for viewing reports, and the specific time of completing health tasks (such as recording diet and checking in for exercise). Wearable device data may include, but is not limited to: the time distribution of daily steps, heart rate change trends, sleep start and end times, and the time points when resting and active states switch. Health intervention response records may include, but are not limited to: the click time after receiving historical push notifications and the time window for completing suggested actions. Life event logs may include, but are not limited to: manually recorded meal times, medication times, and exercise start times.

[0077] Among them, the attribute description information refers to the relatively stable basic demographic, physiological or social role characteristics of the target user that do not change frequently over time. It is used to characterize the target user's identity and long-term life background. For details, please refer to the aforementioned embodiments, which will not be repeated here.

[0078] Frequency statistics and time-segment aggregation refer to the statistical analysis of time-series behavioral information along the time dimension to identify the high-frequency activity periods of target users. For example, frequency statistics could involve calculating the number of times a target user is active in each hour (or at a more granular level) within a 24-hour day. For instance, in the past 30 days, a target user opened a health and wellness app 25 times between 8 PM and 10 PM and 3 times between 9 AM and 11 AM. Time-segment aggregation could involve merging consecutive hours of high activity into candidate push notification periods. For example, if a target user is frequently active between 7 PM and 9:30 PM, then 7 PM to 9:30 PM would be aggregated into a single candidate push notification period.

[0079] The candidate push time periods may be influenced by user attribute information. Therefore, attribute description information can be introduced as prior knowledge, and an attention mechanism can be used to weight and adjust each candidate time period to improve long-term rationality. Specifically, attribute description information is used as a weight adjustment factor, and the attention mechanism is used to correct the candidate push time periods to obtain the target push time period corresponding to the target user. Attribute description information determines which time periods are more consistent with the target user's life role. For example, "night shift nurses" are often active in the early morning, while ordinary white-collar workers are active during the day. The attention mechanism is that the model automatically learns the influence strength of different attributes on different time periods. For example, when "working as a programmer" and "30 years old" are detected, the model can assign a higher attention score to the evening time period; if "70 years old" is detected, a higher attention score may be assigned to the morning or forenoon time period.

[0080] Different users correspond to different push notification time slots. The target push notification time is a specific point in time within the target push notification time slot. For example, a push notification time slot can be randomly selected from the target push notification time slot. Another example is determining the target user's more frequent active time slots from time-series behavior information; if these more frequent active time slots fall within the target push notification time slot, then those more frequent active time slots are used as the target push notification time slots. Yet another example is avoiding the very beginning and end of the target push notification time slot to prevent users from being unprepared or having already left. For example, the first 15 minutes of the target push notification time slot and the last 15 minutes of its end.

[0081] The internal structure of the large-scale push timing prediction model is not limited. Below is an example of a large-scale push timing prediction model that includes: The semantic parsing module is used to perform semantic analysis on medical and health profile data to obtain temporal behavioral information and attribute descriptions. For example, it can be implemented based on a rule engine and a pre-trained language model.

[0082] The time-segment generation module is used to perform frequency statistics and time-segment aggregation on temporal behavior information to obtain candidate push time periods for target users. For example, it can be implemented using a Temporal Convolutional Network (TCN) or a Transformer-based time series encoder (Informer, TimeSformer).

[0083] The correction module uses attribute description information as a weight adjustment factor and employs an attention mechanism to correct the candidate push time periods, thereby obtaining the target push time period corresponding to the target user. For example, it can be implemented through a cross-attention network.

[0084] Among them, the topic generation model extends vertically in semantic depth, horizontally in semantic breadth, and preventatively expands in the time dimension; the copywriting generation model generates personalized and style-adapted candidate medical and health recommendation copy through copywriting style information; and the push timing prediction model provides accurate target push time periods. The three are deeply integrated, working together from the three dimensions of "content", "form" and "timing" to build an intelligent push mechanism that truly adapts to target users, thereby greatly improving user engagement and health management experience.

[0085] In one optional embodiment, after recommending the target medical and health content to the target user, behavioral data from each user of the medical and health application can be collected and batch analysis initiated to summarize the patterns of high-click-rate themes and content, as well as the preference differences among different user groups. This allows for continuous optimization of the theme generation model and the content generation model through a data-driven iterative closed loop, achieving continuous self-evolution and improving the intelligence of the medical and health application. The collected behavioral data is information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of this data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. A corresponding operation entry point is provided for users to choose to authorize or refuse.

[0086] Specifically, log information from multiple users, including target users, is obtained. The log information includes: click behavior on the pushed target medical and health recommendation texts, as well as the initial topic, target topic, medical and health profile data, and topic type corresponding to each target medical and health recommendation text. The topic type belongs to one of the following: vertical extension type, horizontal derivation type, and predictive topic type. Based on the click behavior, the click-through rate of each target topic is calculated, and the first target topic with a click-through rate higher than a set click-through rate threshold is selected. The first target topic corresponds to the first medical and health recommendation text. From the log information, the first initial topic, first medical and health profile data, and first topic type corresponding to the first medical and health recommendation text are obtained. A first sample is constructed using the first medical and health profile data and the first initial topic, and the labels for the first sample are constructed using the first target topic and the first topic type. Supervised fine-tuning of the topic generation model is then performed to update the model parameters of the topic generation model. Figure 2 The document demonstrates the process of recording user click behavior, calculating click-through rates, and constructing samples and their labels during the feedback and optimization phase.

[0087] This system can periodically record log information generated by target users during the generation and delivery of information. By analyzing the log information collected in the previous period within the current period, the system can dynamically optimize the topic generation model within the current period. The log information records the entire process for each user, from the moment they begin using the healthcare application, through the collection of initial topics, generation of target topics, generation of candidate healthcare recommendation texts, delivery of target healthcare recommendation texts, and click behavior on those texts. The log information may include, but is not limited to: click behavior on the delivered target healthcare recommendation texts, and the initial topic, target topic, healthcare profile data, and topic type for each target healthcare recommendation text. The topic type can be one of three types: vertical extension, horizontal derivation, or predictive topic. Vertical extension corresponds to semantic depth extension of the initial topic, horizontal derivation corresponds to semantic breadth extension of the initial topic, and predictive topic corresponds to preventative temporal expansion of the initial topic.

[0088] Based on this, the click-through rate of each target topic is calculated, and the first target topic with a click-through rate higher than the set click-through rate threshold is selected from each target topic. For any first target topic, the target medical and health recommendation copy to which the first target topic belongs is called the first medical and health recommendation copy. From the log information, the first initial topic, the first medical and health profile data, and the first topic type corresponding to the first medical and health recommendation copy are obtained.

[0089] The first sample is constructed based on the first medical and health profile data and the first initial topic. The label of the first sample is constructed based on the first target topic and the first topic type. The topic generation model is then supervised and fine-tuned using the first sample and its label to update the model parameters of the topic generation model.

[0090] Optionally, the first target topic corresponds to one or more first medical and health recommendation texts. Common user features are extracted from the one or more first medical and health recommendation texts, and the first sample is constructed using the common user features and the first initial topic.

[0091] Optionally, the log information also includes: copywriting style information corresponding to each target medical and health recommendation copy; obtaining the first copywriting style information corresponding to the first medical and health recommendation copy from the log information; constructing a second sample using the first medical and health profile data, the first target topic, and the first copywriting style information; constructing the label of the second sample using the first medical and health recommendation copy; and performing supervised fine-tuning of the copywriting generation model to update the model parameters of the copywriting generation model.

[0092] Optionally, the log information also includes: the click time points of each user's click behavior; the click time point corresponding to the first medical and health recommendation copy is obtained from the log information; a third sample is constructed using the first medical and health profile data, and the labels of the third sample are constructed using the click time points to fine-tune the push timing model and update the model parameters of the push timing model.

[0093] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 101 to 104 can be device A; or the execution subject of steps 101 and 102 can be device A, and the execution subject of step 103 can be device B; and so on.

[0094] Furthermore, some processes described in the above embodiments and accompanying drawings include multiple operations appearing in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0095] In this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can also mean indirect receiving and sending. For example, A receiving data sent by B can be understood as A directly receiving data sent by B, or it can be understood as A indirectly receiving data sent by B through other entities such as C. Similarly, B sending data to A can be understood as B sending data directly to A, or it can be understood as B indirectly sending data to A through other entities such as C. Here, C can be one entity, or it can be two or more entities.

[0096] Figure 3 A schematic diagram of the structure of an information recommendation device provided for exemplary embodiments of this specification, such as... Figure 3 As shown, the device includes: a construction module, a topic generation module, a copywriting generation module, and a recommendation module; Module 31 is used to construct medical and health profile data of the target user based on multi-source behavioral data and health record data of the target user in medical and health applications. The medical and health profile data includes at least the initial topics and copywriting style information that the target user is interested in. The topic generation module 32 is used to utilize a large topic generation model, based on medical and health profile data, to semantically expand the initial topic in order to obtain multiple target topics; The copy generation module 33 is used to generate candidate medical and health recommendation copy for multiple target topics based on the copy generation model and copy style information. The recommendation module 34 is used to select the target medical and health recommendation copy from the candidate medical and health recommendation copy and recommend the target medical and health recommendation copy to the target user.

[0097] For detailed descriptions of the implementation methods and effects of the above-mentioned device, please refer to the foregoing embodiments, which will not be repeated here.

[0098] Figure 4 This specification illustrates a schematic diagram of an electronic device provided in an exemplary embodiment, which is applicable to the model training method provided in the foregoing embodiments. For example... Figure 4 As shown, the electronic device 700 mainly consists of a communication interface 702, a user interface 704, a processor 706, and a memory 708. These components are interconnected and communicate with each other through a system bus, network, or other connection mechanism 410. The communication interface 702 enables the device 700 to communicate with other devices, access networks, and transmission networks via analog or digital modulation. For example, the communication interface 702 may include a chipset and antenna for wireless communication with a radio access network or access point. Furthermore, the communication interface 702 can also be a wired interface such as Ethernet, Token Ring, or a USB port, or a wireless interface such as Wi-Fi (Wireless Fidelity), Bluetooth, Global Positioning System (GPS), or wide-area wireless interface such as WiMAX (Wireless Maximum) or LTE (Long Term Evolution). Of course, the communication interface 702 can also support other forms of physical layer interfaces and standard or proprietary communication protocols. The communication interface 702 may also include multiple physical communication interfaces, such as a Wi-Fi interface, a Bluetooth interface, and a wide-area wireless interface.

[0099] User interface 704 includes receiving user input and providing output to the user. Therefore, user interface 704 may include input components such as a keypad, keyboard, touch-sensitive or presence-sensitive panel, computer mouse, trackball, joystick, microphone, still camera, and video camera, and output components such as a display screen (which may be combined with a touch-sensitive panel), CRT (Cathode Ray Tube), LCD (Liquid Crystal Display), LED (Light Emitting Diode), display using DLP (Digital Light Processing) technology, printer, and other known or future similar devices. User interface 704 may also generate auditory output via speakers, speaker jacks, audio output ports, audio output devices, headphones, and other known or future similar devices. In some embodiments, user interface 704 may include software, circuitry, or other forms of logic capable of transmitting and receiving data from external user input / output devices. Additionally or alternatively, electronic device 700 may support remote access from other devices via communication interface 702 or another physical interface (not shown). User interface 704 can be configured to receive user input, the position and movement of which can be indicated by an indicator or cursor described herein. User interface 704 can also be configured as a display device for rendering or displaying text fragments.

[0100] Processor 706 may include one or more general-purpose processors and / or special-purpose processors. Memory 708 may include one or more volatile and / or non-volatile memory components and may be integrated wholly or partially with processor 706. Memory 708 may include removable and non-removable components.

[0101] The processor 706 is capable of executing program instructions 718 (e.g., compiled or uncompiled program logic and / or machine code) stored in memory 708 to perform the various functions described herein.

[0102] Memory 708 may contain non-transitory computer-readable media, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Memory 708 stores program instructions that, when executed by device 700, enable device 700 to perform any of the methods, processes, or functions disclosed in this specification and / or the accompanying drawings. Processor 706 executing program instructions 718 may cause processor 706 to use data 712.

[0103] For example, program instructions 718 may include an operating system 722 (e.g., an operating system kernel, device drivers, and / or other modules) installed on device 700 and one or more applications 720 (e.g., a browser, social application, or game application). Similarly, data 712 may include operating system data 716 and application data 714. Operating system data 716 is primarily accessible to the operating system 722, while application data 714 is primarily accessible to one or more applications 720. Application data 714 may reside in a file system visible or hidden from the user of device 700.

[0104] Application 720 can communicate with operating system 722 through one or more application programming interfaces (APIs). These APIs help application 720 read and / or write application data 714, transmit or receive information via communication interface 702, receive or display information on user interface 704, etc.

[0105] In some terminology, application 720 may be simply referred to as "app". Furthermore, application 720 can be downloaded to device 700 through one or more online app stores or app markets. However, applications can also be installed on device 700 in other ways, such as through a web browser or a physical interface on electronic device 700 (e.g., a USB port).

[0106] Accordingly, embodiments of this specification also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, embodiments of this specification also provide a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above-described method embodiments. It should be understood that each step or combination of steps in the above-described method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above-described method embodiments.

[0107] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, product, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes that element.

[0108] This specification uses specific terms to describe embodiments thereof. Terms such as "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of those different embodiments or examples, without contradiction.

[0109] The terminology used in the embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “the,” and “the” used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. “Multiple” generally includes at least two, but does not exclude the inclusion of at least one. “A plurality” generally includes at least two, but does not exclude the inclusion of at least one.

[0110] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0111] The above are merely embodiments of this specification and are not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. An information recommendation method, characterized in that, include: Based on multi-source behavioral data and health record data of target users in medical and health applications, a medical and health profile data of the target users is constructed. The medical and health profile data includes at least the initial topics and copywriting style information that the target users are interested in. Using a topic-based large model, based on the medical and health profile data, semantic expansion is performed on the initial topic to obtain multiple target topics; Using a copywriting generation model, candidate medical and health recommendation copy is generated for each of the multiple target topics based on the copywriting style information. Select a target medical and health recommendation from the candidate medical and health recommendation texts, and recommend the target medical and health recommendation text to the target user; The topic generation model corresponds to a medical and health knowledge graph, which includes multiple topic nodes. These topic nodes have vertical hierarchical relationships, horizontal hierarchical relationships, and temporal relationships. Using a topic-based large-scale model, based on the aforementioned medical and health profile data, semantic expansion is performed on the initial topic to obtain multiple target topics, including: Semantic parsing is performed on the medical and health profile data to obtain the parsing results, which include at least the initial topic; using the initial topic as the query benchmark, node traversal is performed in the medical and health knowledge graph to locate the first topic node corresponding to the initial topic; Using the first topic node as the anchor point, queries are performed in the medical and health knowledge graph along the vertical hierarchical relationship, horizontal hierarchical relationship, and temporal relationship of the first topic to obtain multiple related topic nodes; The topic generation model is invoked, and multiple target topics are generated based on the topic tags and content summaries of the multiple associated topic nodes.

2. The method according to claim 1, characterized in that, Selecting a target medical and health recommendation from the candidate medical and health recommendation texts includes: Add the candidate medical and health recommendation texts corresponding to each of the multiple target topics to the target buffer pool corresponding to the target user, wherein different users correspond to different buffer pools; During the target push period, select target medical and health recommendation text from the target buffer pool.

3. The method according to claim 1, characterized in that, The parsing results also include topic context information corresponding to the initial topic, as well as steady-state feature information and environmental context information corresponding to the target user; Using the first topic node as an anchor point, queries are performed in the medical and health knowledge graph along the vertical hierarchical relationship, horizontal hierarchical relationship, and temporal relationship of the first topic to obtain multiple related topic nodes, including: Based on the topic context information, in the medical and health knowledge graph, the child nodes corresponding to the first topic node are queried as associated topic nodes along the vertical hierarchical relationship. Based on the steady-state feature information, in the medical and health knowledge graph, the sibling nodes that share the same parent node with the first topic node are queried along the horizontal hierarchical relationship and used as associated topic nodes. Based on the environmental context information, in the medical and health knowledge graph, other topic nodes that have a causal relationship with the first topic node are queried along the temporal correlation and used as associated topic nodes.

4. The method according to claim 1 or 2, characterized in that, Using a copywriting generation model, candidate medical and health recommendation copy is generated for each of the multiple target topics based on the copywriting style information, including: The medical and health profile data and the multiple target topics are input into the large-scale copywriting generation model. For any target topic, the user context information associated with the target topic is parsed from the medical and health profile data; Using the aforementioned copywriting style information as a constraint, semantic expansion and content generation are performed on the target topic to obtain push copy; Based on the push notification and the user context information, medical and health query information is generated. This medical and health query information is used to input into the medical and health question and answer model associated with the target application, so that the medical and health question and answer model can generate answers. The push notification text and the medical and health query information are used as candidate medical and health recommendation texts.

5. The method according to claim 4, characterized in that, Recommending the target medical and health recommendation text to the target user includes: providing the target push text included in the target medical and health recommendation text to the target user; The method further includes: In response to a user's action on the target push notification, the target medical and health query information in the target medical and health recommendation text is input into the medical and health question-and-answer model to generate an answer, thereby obtaining medical and health answer information; The medical and health answers will be displayed on the question and answer page of the target application.

6. The method according to claim 2, characterized in that, Also includes: The medical and health profile data is input into the push timing prediction model. Semantic parsing is performed on the medical and health profile data to obtain the target user's temporal behavior information and the target user's attribute description information. Frequency statistics and time period aggregation are performed on the temporal behavior information to obtain candidate push time periods. Using the attribute description information as a weight adjustment factor, the candidate push time period is corrected using an attention mechanism to obtain the target push time period.

7. The method according to claim 1 or 2, characterized in that, After recommending the target medical and health recommendation text to the target user, the process also includes: Obtain log information for multiple users, including the target user. The log information includes: click behavior for the pushed target medical and health recommendation text, as well as the initial topic, target topic, medical and health profile data, and topic type corresponding to each target medical and health recommendation text. Based on the click behavior, the click-through rate of each target topic is calculated, and a first target topic with a click-through rate higher than a set click-through rate threshold is selected; the first target topic corresponds to a first medical and health recommendation text. From the log information, obtain the first initial topic, the first medical and health profile data, and the first topic type corresponding to the first medical and health recommendation text; A first sample is constructed using the first medical and health profile data and the first initial topic. The labels of the first sample are constructed using the first target topic and the first topic type. The topic generation model is then subjected to supervised fine-tuning to update the model parameters of the topic generation model.

8. The method according to claim 7, characterized in that, The log information also includes: the copywriting style information corresponding to each target medical and health recommendation copywriting; From the log information, obtain the first copywriting style information corresponding to the first medical and health recommendation copywriting; A second sample is constructed using the first medical and health profile data, the first target topic, and the first copywriting style information. The tags of the second sample are constructed using the first medical and health recommendation copywriting. The copywriting generation model is then subjected to supervised fine-tuning to update the model parameters of the copywriting generation model.

9. The method according to claim 7, characterized in that, The log information also includes: the click time points of each of the multiple users for the click behavior; From the log information, obtain the click time point corresponding to the first medical and health recommendation text; A third sample is constructed using the first medical and health profile data, and the labels of the third sample are constructed using the click time points. The push timing model is then fine-tuned to update the model parameters of the push timing model.

10. An information recommendation device, characterized in that, include: Modules include: building module, theme generation module, copywriting generation module, and recommendation module. The construction module is used to construct the target user's medical and health profile data based on the target user's multi-source behavioral data and health record data in medical and health applications. The medical and health profile data includes at least the initial topics and copywriting style information that the target user is interested in. The topic generation module is used to utilize a large topic generation model to semantically expand the initial topic based on the medical and health profile data, so as to obtain multiple target topics. The copywriting generation module is used to generate candidate medical and health recommendation copywriting corresponding to each of the multiple target topics based on the copywriting style information using a copywriting generation model. The recommendation module is used to select a target medical and health recommendation text from the candidate medical and health recommendation texts, and recommend the target medical and health recommendation text to the target user; The topic generation model corresponds to a medical and health knowledge graph, which includes multiple topic nodes. These topic nodes have vertical hierarchical relationships, horizontal hierarchical relationships, and temporal relationships. When the topic generation module utilizes the topic generation model and, based on the medical and health profile data, semantically expands the initial topic to obtain multiple target topics, it is specifically used for: Semantic parsing is performed on the medical and health profile data to obtain the parsing results, which include at least the initial topic; using the initial topic as the query benchmark, node traversal is performed in the medical and health knowledge graph to locate the first topic node corresponding to the initial topic; Using the first topic node as the anchor point, queries are performed in the medical and health knowledge graph along the vertical hierarchical relationship, horizontal hierarchical relationship, and temporal relationship of the first topic to obtain multiple related topic nodes; The topic generation model is invoked, and multiple target topics are generated based on the topic tags and content summaries of the multiple associated topic nodes.

11. An electronic device, characterized in that, include: A memory and a processor; the memory is used to store one or more computer instructions; the processor is used to execute the one or more computer instructions for: performing the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it is able to perform the steps of the method according to any one of claims 1-9.

13. A computer program product, characterized in that, include: A computer program or instructions that, when executed by a processor, can perform the steps of the method according to any one of claims 1-9.

Images