Method and device for personalized recommendation of elderly care services

Abstract

The application provides an old-age care service personalized recommendation method and device, relates to the technical field of service recommendation, and comprises the following steps: obtaining personal portrait information of a target old person, fusing the personal portrait information with an old-age care knowledge graph, and constructing an old person's preferred old-age care knowledge graph; performing multi-round denoising on the old person's preferred old-age care knowledge graph to obtain a target old-age care knowledge graph; based on the target old-age care knowledge graph and old person-care service interaction data, performing joint training through a multi-task learning model to output a final embedding representation of the target old person and a final embedding representation of each candidate old-age care service; calculating the matching degrees between the final embedding representation of the target old person and the final embedding representations of the candidate old-age care services respectively, performing matching degree sorting, and outputting a plurality of old-age care service recommendation results corresponding to the target old person. The application realizes accurate recommendation of old-age care services.

Description

Technical Field

[0001] This invention relates to the field of service recommendation technology, and in particular to a method and apparatus for personalized recommendation of elderly care services. Background Technology

[0002] With the continuous improvement of the smart elderly care system, personalized recommendations for elderly care have become a key link in optimizing nursing service processes and improving the elderly's experience. This covers diverse scenarios such as matching caregivers, recommending nursing programs, fitting assistive devices, and customizing rehabilitation plans. Currently, knowledge graph (KG)-based elderly care recommendation methods effectively alleviate the problems of sparse elderly-nursing service interaction data and insufficient coverage of long-tail nursing programs faced by traditional collaborative filtering methods by integrating external elderly care knowledge such as disease-nursing programs, assistive devices-applicable populations, and caregivers-specialty areas. This provides technical support for accurately matching the needs of the elderly with elderly care resources.

[0003] However, in real-world elderly care scenarios, existing knowledge-based (KG) recommendation methods still have significant limitations, failing to meet the requirements of professionalism and safety in elderly care. A prime example is the severe noise interference in the elderly care KG. The KG typically contains a large amount of redundant information irrelevant to the recommendation task, such as non-core nursing equipment parameters, redundant industry literature citations, and weakly correlated collaborative relationships among elderly care institutions. This noise leads to an overemphasis on elderly-item (entity) interaction signals during model training, preventing the effective encoding of core elderly care knowledge into the embedded representations of users and services. Furthermore, indiscriminately introducing low-relevance KG information can easily cause "recommendation bias," such as recommending wheelchair-fitting services to elderly individuals with good mobility, or recommending high-cost fall prevention care programs to elderly individuals with low fall risk, resulting in resource waste and insufficient care suitability.

[0004] Therefore, an effective technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and apparatus for personalized recommendation of elderly care services. Through a full-process approach of "knowledge fusion - dual denoising - multi-task optimization," it achieves effective denoising of the elderly care knowledge graph, thereby enabling accurate recommendation of elderly care services.

[0006] In a first aspect, the present invention provides a method for personalized recommendation of elderly care services, the method comprising the following steps: Obtain personal profile information of the target elderly person; the personal profile information includes at least one of the following: basic information, health data, and care preferences. By integrating the personal profile information of the target elderly with the elderly care knowledge graph, an elderly preference elderly care knowledge graph representing the personalized care needs of the target elderly is constructed. The elderly care preference knowledge graph is subjected to multiple rounds of noise reduction to obtain the target elderly care knowledge graph. Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output the elderly care service recommendation results for the first preset number of target elderly persons according to the ranking of the matching degrees.

[0007] According to the present invention, a personalized recommendation method for elderly care services includes performing multi-round denoising on the elderly's preferred elderly care knowledge graph to obtain a target elderly care knowledge graph, comprising: The first round of noise trimming is performed on the elderly preference elderly care knowledge graph to remove redundant edges, resulting in a preliminary denoised elderly care knowledge graph. A second round of knowledge denoising is performed on the initially denoised elderly care knowledge graph. Based on the statistical independence criterion, redundant information that is irrelevant to the current care needs of the target elderly is filtered out to obtain the target elderly care knowledge graph.

[0008] According to a personalized recommendation method for elderly care services provided by the present invention, the first round of noise trimming of the elderly preference elderly care knowledge graph, which removes redundant edges from the elderly preference elderly care knowledge graph to obtain a preliminarily denoised elderly care knowledge graph, includes: The importance score of each edge in the knowledge graph of elderly preferences for elderly care is calculated. The importance score is obtained through multi-dimensional evaluation based on entity relevance, elderly health matching degree, and elderly care fit degree. The entity relevance is determined by the semantic distance between entities in the knowledge graph of elderly preferences for elderly care. The elderly health matching degree is determined by the correlation between the disease diagnosis in the target elderly’s electronic medical record and the elderly care entity. The elderly care fit degree is determined by the degree of fit between the elderly’s living habits, living environment and care entity. Based on the edge importance score corresponding to each edge, edges with an edge importance score lower than the dynamic threshold are identified as redundant edges and removed to obtain the preliminary denoised elderly care knowledge graph.

[0009] According to a personalized recommendation method for elderly care services provided by the present invention, the step of fusing the personal profile information of the target elderly person with an elderly care knowledge graph to construct an elderly care knowledge graph representing the personalized care needs of the target elderly person includes: Identify the personal profile information of the target elderly person and the multi-hop paths between entities in the elderly care knowledge graph; If any target entity among the entities can be associated through the multi-hop path and the matching degree between the elderly node and the target entity meets the preset matching degree threshold, a direct association edge is established between the elderly node and the target entity; the elderly node is constructed based on the personal profile information of the target elderly. Based on the directly related edges and the elderly care knowledge graph, the elderly preference elderly care knowledge graph is constructed.

[0010] According to a personalized recommendation method for elderly care services provided by the present invention, the step of fusing the personal profile information of the target elderly person with an elderly care knowledge graph to construct an elderly care knowledge graph representing the personalized care needs of the target elderly person includes: Based on the personal profile information of the target elderly, structured information including a risk assessment report, a vital sign monitoring list, and an initial care item list is generated; The structured information is fused with the elderly care knowledge graph to obtain the elderly preference elderly care knowledge graph.

[0011] According to the present invention, a personalized recommendation method for elderly care services, wherein the method is based on the target elderly care knowledge graph and elderly-care service interaction data, and is jointly trained through a multi-task learning model to output the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, including: Based on the target elderly care knowledge graph and the elderly-care service interaction data, a nursing interaction view and an elderly preference elderly care view are constructed. Graph attention networks are then trained on the nursing interaction view and the elderly preference elderly care view to obtain nursing interaction embedding representations and elderly care knowledge embedding representations, respectively. The nursing interaction embedding representation and the elderly care knowledge embedding representation are mapped to the same semantic space and compared and aligned to enhance the similarity between positive sample pairs and reduce the distance between negative sample pairs, resulting in aligned knowledge embeddings. The definitions of the positive and negative sample pairs conform to the constraints of the elderly care scenario. Positive sample pairs include at least one of the following: nursing interaction embedding and elderly care knowledge embedding for the same elderly person; nursing project embedding and nursing staff expertise embedding for the same disease; nursing needs embedding and nursing project embedding for the same elderly person; rehabilitation nursing embedding and nursing staff expertise embedding for the same underlying disease. Negative sample pairs include at least one of the following: elderly person embedding and non-health-related elderly care entity embedding; elderly care entity embedding and non-fitted nursing entity embedding for different diseases; elderly person embedding and wheelchair-fitted service embedding for elderly people with mobility impairments. Based on the aligned knowledge embeddings corresponding to the target elderly care knowledge graph and the elderly-care service interaction data, and using multi-task loss and elderly care compliance constraints, the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service are output through joint training with an adaptive momentum optimizer.

[0012] According to the present invention, a personalized recommendation method for elderly care services includes a multi-task loss comprising a personalized recommendation loss for elderly care, a comparison and alignment loss for elderly care, and a bottleneck denoising loss for elderly care.

[0013] According to a personalized recommendation method for elderly care services provided by the present invention, the step of outputting multiple elderly care service recommendation results corresponding to the target elderly person based on the ranking of the matching degrees includes: The system outputs multiple elderly care service recommendations, corresponding personalized health advice, and traceable elderly care recommendation descriptions, ordered in descending order of matching degree. The multiple elderly care service recommendations for the target elderly person comply with elderly care standards.

[0014] According to a personalized recommendation method for elderly care services provided by the present invention, the elderly care knowledge graph is updated through the following steps: The elderly care knowledge graph is updated based on regularly synchronized external data sources. The semantic distances between entities in the updated elderly care knowledge graph are recalculated, and the recalculated semantic distances between entities are updated in the metadata storage layer of the updated elderly care knowledge graph.

[0015] Secondly, the present invention also provides a personalized recommendation device for elderly care services, the device comprising the following modules: The acquisition unit is used to acquire personal profile information of the target elderly person; the personal profile information includes at least one of basic information, health data, and care preferences. The knowledge fusion unit is used to fuse the personal profile information of the target elderly with the elderly care knowledge graph to construct an elderly preference elderly care knowledge graph that represents the personalized care needs of the target elderly. A multi-round denoising unit is used to perform multi-round denoising on the elderly preference elderly care knowledge graph to obtain the target elderly care knowledge graph. The multi-task optimization unit is used to jointly train a multi-task learning model based on the knowledge graph of the target elderly care and the elderly-care service interaction data, and output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. A personalized recommendation unit is used to calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

[0016] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the personalized recommendation method for elderly care services as described above.

[0017] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the personalized recommendation method for elderly care services as described above.

[0018] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the personalized recommendation method for elderly care services as described above.

[0019] This invention provides a method and apparatus for personalized recommendation of elderly care services. First, it acquires the personal profile information of a target elderly person, which includes at least one of basic information, health data, and care preferences. Then, it fuses the target elderly person's personal profile information with an elderly care knowledge graph to construct an elderly person preference elderly care knowledge graph representing the target elderly person's personalized care needs. Further, it performs multiple rounds of denoising on the elderly person preference elderly care knowledge graph to obtain a target elderly care knowledge graph. Next, based on the target elderly care knowledge graph and elderly-care service interaction data, it performs joint training through a multi-task learning model to output the final embedding representation of the target elderly person and the final embedding representations of each candidate elderly care service. Finally, it calculates the matching degree between the final embedding representation of the target elderly person and the final embedding representations of each candidate elderly care service, and outputs the elderly care service recommendation results for the top preset number of target elderly people according to the ranking of the matching degrees. This invention achieves effective denoising of the elderly care knowledge graph through a complete process of "knowledge fusion - dual denoising - multi-task optimization," thereby achieving accurate recommendation of elderly care services. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating existing methods for recommending elderly care services.

[0022] Figure 2 This is one of the flowcharts illustrating the personalized recommendation method for elderly care services provided by this invention.

[0023] Figure 3 This is the second flowchart illustrating the personalized recommendation method for elderly care services provided by this invention.

[0024] Figure 4 This is a schematic diagram of the personalized elderly care service recommendation device provided by the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first node can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0028] To more clearly understand the various embodiments provided by the present invention, the prior art is first described as follows: Figure 1 This is a flowchart illustrating existing technologies for recommending elderly care services, such as... Figure 1 As shown, the method includes the following: The system inputs a medical knowledge graph; a static noise filtering module removes redundant edges based on a fixed threshold; inputs patient-medical service interaction data; performs simple fusion of knowledge and interaction features without balancing the weights of knowledge and interaction signals; filters redundancy based on the fused single graph; verifies surface rules such as drug allergies, without regional medical classification adaptation; calculates the matching degree and outputs Top-K (top K) recommendation results based on the matching degree.

[0029] Existing knowledge graph-based recommendation methods still have significant limitations and are difficult to meet the professionalism and safety requirements of elderly care: First, the knowledge graph for elderly care suffers from severe noise interference.

[0030] Secondly, the fragmented and dynamically adapted needs profiles of the elderly are a significant issue. Elderly individuals often suffer from multiple coexisting conditions, require care across different institutions, and experience dynamic changes in their care stages, resulting in fragmented needs profiles. Existing methods often employ static Key-Genius (KG) fusion strategies, which struggle to capture the dynamic adaptation relationship between "health status and care services." For example, they cannot prioritize follow-up care services for elderly individuals in the stable phase of chronic diseases, or delay recommending rehabilitation training programs for elderly individuals in the postoperative recovery phase, impacting the timeliness and relevance of recommendations.

[0031] Furthermore, there is a lack of compliance constraints and low knowledge utilization in elderly care. Recommendations for elderly care services must strictly adhere to elderly care standards, but existing methods lack specific compliance verification mechanisms for elderly care scenarios, making non-compliant recommendations prone to occur. Meanwhile, related research confirms that existing KG-driven recommendation methods are inefficient at encoding elderly care knowledge. For example, in some models, knowledge entities strongly correlated with the elderly's health status receive extremely low attention weights, resulting in the ineffective utilization of core elderly care knowledge and further reducing recommendation accuracy.

[0032] Finally, the issues of data sparsity and cold start are prominent. The interaction data of long-tail elderly care resources such as primary care services and rare disease-related care projects are extremely sparse, making it difficult for existing methods to mine elderly preferences with limited data. At the same time, cold-start elderly care entities such as newly hired caregivers and newly added rehabilitation care technologies are difficult to effectively recommend to target elderly people due to the lack of historical interaction records, resulting in uneven distribution of elderly care resources.

[0033] While some elderly care recommendation tools have emerged in the market, these tools lack specific mechanisms designed for the noise characteristics, compliance requirements, and dynamic needs of elderly care knowledge bases (KGs). This results in low knowledge utilization, insufficient robustness of recommendations, and risks associated with compatibility. Therefore, there is an urgent need for a recommendation framework that can proactively identify noise in elderly care knowledge bases, align elderly health conditions with elderly care knowledge, and meet elderly care compliance requirements. This framework would enable "safe, accurate, and dynamic" matching of elderly care services, promoting the practical application of smart elderly care recommendation technology.

[0034] To address the problems existing in the prior art, this invention provides a personalized recommendation method for elderly care based on knowledge denoising. Through a complete process of "knowledge fusion - noise trimming - dual-view alignment - bottleneck denoising - multi-task optimization", it achieves denoising and accurate recommendation of elderly care knowledge graph, while meeting the compliance requirements of elderly care.

[0035] The following is combined with Figures 2 to 5 The present invention describes a method and apparatus for personalized recommendation of elderly care services.

[0036] Figure 2 This is one of the flowcharts illustrating the personalized recommendation method for elderly care services provided by this invention, such as... Figure 2 As shown, the method includes the following: Step 201: Obtain the personal profile information of the target elderly person; the personal profile information includes at least one of the following: basic information, health data, and care preferences.

[0037] The subject executing the personalized recommendation method for elderly care services provided by this invention can be an electronic device, or any other subject capable of implementing the personalized recommendation method for elderly care services, such as other personalized recommendation systems for elderly care services.

[0038] Among them, the target elderly are those who need to be recommended for elderly care services. "Elderly care services" refers to a series of medical assistance, daily care, rehabilitation training and safety protection programs provided by institutions, communities or families to maintain or improve the physical function and quality of life of the elderly (especially the disabled, semi-disabled, chronically ill and very old).

[0039] Elderly care services can be divided into the following four categories: 1. Nursing and Medical Assistance Programs: This category includes "Nursing Programs," "Diagnosis and Treatment Programs," and "Rehabilitation Plans." Examples are as follows: Basic care: "Turning over care" program recommended for disabled elderly.

[0040] Medical care: "Regular blood glucose monitoring" or "insulin injection guidance" are recommended for elderly people with diabetes.

[0041] Postoperative rehabilitation: For elderly patients who have undergone hip replacement surgery, a "rehabilitation training program" or "physical therapy" is recommended.

[0042] 2. Personnel matching services: These services refer to recommending suitable professionals (such as "nursing staff matching", "doctor's expertise", "family doctor") to the elderly.

[0043] For example, for elderly people with Parkinson's disease, a dedicated caregiver with neurological nursing experience is recommended. For elderly people in areas with a high incidence of endemic diseases, a local endemic disease specialist or a specialist at a regional medical center is recommended.

[0044] 3. Assistive Devices and Age-Friendly Products: This category refers to devices that help seniors regain their independent living abilities. For example, "Assistive Device Fitting." Examples are as follows: Mobility assistance: For elderly people who have difficulty moving but can still walk, we recommend "four-legged canes" or "walking aids"; for elderly people who are completely unable to walk, we recommend "wheelchairs".

[0045] Safety precautions: We recommend "fall monitoring devices" or "bathroom handrail installation services" for elderly people at high risk of falling.

[0046] Daily living assistance: Recommend hearing aids for elderly people with hearing loss.

[0047] 4. Health advice and risk control: "Personalized health advice", "risk assessment report", "emergency services", etc., focusing on prevention and management.

[0048] In practical applications, the first step is to obtain the target elderly person's personal profile information, such as basic information, health data, and care preferences. This personal profile information includes, but is not limited to: basic demographic information (age, gender, place of residence), health data (history of underlying diseases, allergies, current medication use, physical examination indicators), physical function status (scores for activities of daily living, fall risk level, cognitive function assessment results), and care preferences (expected service time, service type preference, language preference). Optionally, the raw data can be preprocessed to ensure it conforms to the input specifications of subsequent modules, thus obtaining the target elderly person's personal profile information.

[0049] Step 202: Integrate the personal profile information of the target elderly with the elderly care knowledge graph to construct an elderly care knowledge graph that represents the personalized care needs of the target elderly.

[0050] Specifically, the elderly care knowledge graph can be understood as a structured knowledge network that integrates multi-source data, involves domain experts in ontology design, extracts knowledge using deep learning technology, and stores it in a graph database. It covers core entities such as diseases, symptoms, medications, care programs, caregivers, and assistive devices, along with their complex relationships, providing a foundational knowledge base for the subsequent integration and construction of an "elderly preference elderly care knowledge graph."

[0051] The pre-construction process of the elderly care knowledge graph includes the following steps 1-4: Step 1: Determine the data sources for the knowledge graph This includes authoritative knowledge bases and clinical guidelines, clinical diagnosis and treatment data, care plans and nursing records, as well as obtaining the latest medical literature, drug catalogs, and medical device standards through web crawling technology.

[0052] Step 2, Body Layer Design Define entity types (e.g., elderly, disease, symptoms, medicine, care items, assistive devices, caregivers); Types of relationships between entities (e.g., complications, contraindications, treatment methods, nursing staff expertise).

[0053] Step 3: Knowledge Extraction and Fusion (Data Layer Construction) Transform unstructured or semi-structured data into structured triples (including entity recognition, relation extraction, and multi-hop path discovery).

[0054] Step 4: Knowledge Storage and Representation Graph database storage: Import the extracted "entity-relationship-entity" triple data into the graph database for storage and visualization.

[0055] Semantic distance calculation: After the graph is built, the semantic distance between entities needs to be pre-calculated so that subsequent modules (such as the noise trimming module) can evaluate the importance of edges.

[0056] In practical applications, the personal profile information of the target elderly person is integrated with the elderly care knowledge graph to construct an elderly care knowledge graph that represents the personalized care needs of the target elderly person. Specifically, the personal profile information of the target elderly person is used as the elderly person node.

[0057] The following is a specific implementation of a method for constructing an elderly care preference knowledge graph based on multi-hop paths: integrating the personal profile information of the target elderly with the elderly care knowledge graph to construct an elderly care preference knowledge graph that represents the personalized care needs of the target elderly.

[0058] Step 1: Physical mapping of elderly person's portrait information First, the target elderly person's personal profile information obtained through the elderly care personal profile input module is standardized and linked to entities. The personal profile information includes, but is not limited to: basic demographic information (age, gender, place of residence), health data (history of underlying diseases, history of allergies, current medication status, physical examination indicators), physical function status (score of activities of daily living, fall risk level, cognitive function assessment results), and care preferences (expected service time, service type preference, language preference).

[0059] Specifically, an entity linking method based on dictionary and rule matching is adopted to map key fields in the profile information (such as "type 2 diabetes", "grade III hypertension", "post-left knee replacement surgery") to corresponding entity nodes in the elderly care knowledge graph. If no completely matching entity exists in the graph, a new temporary entity node is created in the graph and its attribute information is labeled, awaiting subsequent fusion and verification.

[0060] Step 2: Discovery of multi-hop paths and calculation of association strength After completing the entity mapping of the portrait information, the system calls the path discovery engine in the elderly care knowledge fusion module to execute a meta-path-based random walk algorithm in the elderly care knowledge graph to identify the multi-hop paths between the entities (source entities) associated with the target elderly and each candidate nursing service entity (target entity).

[0061] As an optional implementation, the multi-hop path includes at least the following two typical path patterns: Pathway Pattern A (Medical Pathway): Elderly person → Disease → Treatment guidelines → Recommended examination items → Qualified doctor / examination institution; Pathway Model B (Nursing Pathway): Elderly person → Physical dysfunction → Adaptation of assistive devices → Standard nursing procedures → Professional nursing staff.

[0062] For each discovered multi-hop path, the system calculates its path confidence score. The confidence score is calculated as the product of the confidence scores of each hop relationship along the path. For example, if the confidence score of the "elderly → disease" edge is 0.95, the confidence score of the "disease → treatment guidelines" edge is 0.90, and the confidence score of the "guidelines → recommended examination items" edge is 0.85, then the confidence score of the entire path is approximately 0.95 × 0.90 × 0.85 ≈ 0.73.

[0063] Step 3: Establish direct associations based on matching degree thresholds The elderly care knowledge fusion module determines whether to establish a direct association edge between the target elderly node and the target care entity node based on the path confidence calculated in step 2. Specifically: If the target elderly person is associated with the target care entity through any multi-hop path, and the confidence of that path is greater than a preset first threshold (e.g., 0.7), then a "potential association" edge is established between the elderly person node and the target entity node, and the initial weight of the edge is the confidence of that path; if the target elderly person is associated with the same target entity through multiple different multi-hop paths, then the maximum value or weighted average of the confidence of all paths is taken as the final weight of the association edge, and all supporting paths are retained as traceable information for recommendation.

[0064] Step 4: Multi-dimensional dynamic adjustment of fusion weights After establishing directly related edges, the elderly care knowledge fusion module further dynamically adjusts the weights of the edges by combining elderly profile information. The adjustment factors include, but are not limited to: (1) Interaction frequency factor: If the target elderly person has received similar care services to this entity in the past, the weight will be positively enhanced according to the interaction frequency. The higher the frequency, the greater the enhancement. (2) Health-related factors: Based on the current health status (such as acute phase, recovery phase, and stable phase) in the elderly’s electronic medical records, the degree of matching between the nursing entity and the current stage is calculated. The higher the degree of matching, the greater the weight. (3) Urgency factor: If the elderly are currently at risk of emergency (such as high risk of falling or high risk of pressure sores), then the nursing entities related to the prevention and control of such risks will be given a higher urgency weight.

[0065] Finally, the weighted edges were officially added to the elderly preference elderly care knowledge graph, forming a subgraph structure that represents the personalized care needs of the target elderly.

[0066] Optionally, the integration of elderly care knowledge can dynamically adjust the "health relevance" weight based on the elderly's historical care effect data to further strengthen the association priority of high-value elderly care entities; an elderly privacy protection mechanism can be introduced to de-identify sensitive information in electronic medical records to ensure compliant data use; for elderly people with duplicate medical records, the association edge weight between the elderly person and "disease-attending physician" and "disease-long-term follow-up project" can be strengthened, and the weight adjustment can be dynamically determined based on the entity coding score; for elderly people with long-term care needs, the association edge weight between them and "basic disease-dedicated caregiver" and "risk level-prevention and control care project" can be strengthened.

[0067] Step 203: Perform multiple rounds of noise reduction on the elderly's preference knowledge graph for elderly care to obtain the target knowledge graph for elderly care.

[0068] Specifically, after obtaining the knowledge graph of elderly people's preferences for elderly care through knowledge fusion, the knowledge graph of elderly people's preferences for elderly care is subjected to multiple rounds of noise reduction.

[0069] For example, the process includes two rounds of denoising: the first round (structural denoising) removes general redundant information that is common in the graph and irrelevant to the recommendation task; the second round (semantic denoising), also known as knowledge denoising, filters personalized redundant information that is irrelevant to the current real-time care needs of the elderly. The first round of denoising (noise trimming) focuses on structural optimization at the graph level. By evaluating the semantic relevance, health matching degree, and service suitability of each edge, it dynamically trims low-importance redundant edges, constructing a preliminary denoised graph with a clear structure and preserved core relationships, laying the foundation for subsequent learning. The second round of denoising (information bottleneck) focuses on personalized screening at the semantic level. Based on the statistical independence criterion, it calculates the mutual information between each piece of knowledge in the preliminary denoised graph and the current care needs of the target elderly, retaining only knowledge with mutual information values ​​higher than a dynamic threshold (i.e., strong dependencies), and removing redundant information that is statistically independent of the current needs, ensuring that the knowledge input to the recommendation model is all highly relevant and highly suitable core care information. This invention achieves a gradual purification of elderly care knowledge from "general and clean" to "personalized and precise" through a two-round progressive noise reduction mechanism.

[0070] The two-round denoising process has a clear division of labor and organic connection in terms of technical methods, processing targets, and optimization goals: the first round addresses the problem of "the graph itself being too dirty," while the second round addresses the problem of "irrelevant content still remaining in the clean graph," jointly ensuring that the final denoised graph (target elderly care knowledge graph) not only conforms to nursing standards but also accurately adapts to individual needs. The key differences and technical descriptions of the two rounds of denoising are as follows: The key points of the first round of noise reduction are shown in Table 1 below: Table 1:

[0071] The essence of the first round of denoising is to prune edges that "may appear as noise to anyone," based on the graph structure itself and general rules. It does not rely on the dynamic health status of the elderly, but rather establishes a clean basic graph skeleton.

[0072] The key points of the second round of noise reduction are shown in Table 2 below: Table 2:

[0073] The essence of the second round of noise reduction is to further filter out the "truly useful" knowledge from the "preliminarily clean" graph based on information theory and the real-time needs of the elderly. It achieves a leap from "general clean" to "personalized precision".

[0074] Step 204: Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services.

[0075] Specifically, elderly-nursing service interaction data refers to the historical nursing services corresponding to different elderly people collected through big data. In other words, elderly-nursing service interaction data is used to reflect the correspondence between elderly people and elderly care services.

[0076] The final embedded representation of the target elderly person is a multi-dimensional digital vector. It is not just a simple encoding of the elderly person's profile information, but a comprehensive representation that deeply integrates the elderly person's static attributes (age, medical history), dynamic health status (current risk level), behavioral preferences (historical interactions), and core care needs abstracted from the knowledge graph after model learning.

[0077] The final embedded representation of candidate elderly care services is a digital vector corresponding to each specific service (such as "Dr. Zhang's rehabilitation therapy" or "a certain model of pressure ulcer prevention air mattress"). It integrates the service's functional attributes, required qualifications, applicable population characteristics, and associations with various diseases learned from the knowledge graph.

[0078] In practical applications, the target elderly care knowledge graph and elderly-care service interaction data are used as model inputs. The model is jointly trained through a multi-task learning model to output multiple elderly care service recommendations corresponding to the target elderly.

[0079] Step 205: Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output the recommendation results of multiple elderly care services corresponding to the target elderly person according to the ranking of the matching degrees.

[0080] Specifically, the matching degree between the target elderly and each elderly care service is calculated based on the final embedding of the elderly and the final embedding of each candidate elderly care service (the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service respectively), and the Top-K elderly care service recommendation results are output in descending order of matching degree.

[0081] The recommendation logic and performance requirements for the output of elderly care recommendation results include: (1) Matching degree calculation with weighted urgency of elderly care: For elderly care services associated with acute symptoms, the matching degree is multiplied by a specific coefficient to ensure that emergency elderly care services are given priority recommendation; (2) Recommendation performance meets the following requirements: On the dataset of elderly people with chronic diseases, the proportion of elderly care services that match the health status of the elderly in the Top-K recommendations reaches a high proportion; on the dataset of disabled elderly people, the proportion of services that meet the care needs in the Top-K recommendations reaches a high proportion; on the dataset of high-risk elderly people such as those with high risk of falls, the proportion of Top-K recommendations that include risk prevention and control care projects and emergency care services reaches a high proportion, and the recommendation results must be traceable.

[0082] Optionally, the elderly care recommendation output can support the elderly or caregivers to customize recommendation dimensions: preferences such as "priority of medical insurance coverage", "priority of distance", and "priority of positive reviews" can be selected, and the system will dynamically adjust the matching degree calculation weight; the recommendation results can generate a QR code traceability link, and by scanning the code, the knowledge graph entity relationship on which the recommendation is based, the original text of the elderly care standard, and the qualification certificate of the caregiver can be viewed, further improving the credibility of the recommendation.

[0083] The method provided in this invention first obtains the personal profile information of the target elderly person, which includes at least one of basic information, health data, and care preferences. Then, the personal profile information of the target elderly person is fused with an elderly care knowledge graph to construct an elderly care preference knowledge graph representing the personalized care needs of the target elderly person. Further, the elderly preference knowledge graph undergoes multiple rounds of denoising to obtain the target elderly care knowledge graph. Then, based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly person and the final embedding representations of each candidate elderly care service. Finally, the matching degree between the final embedding representation of the target elderly person and the final embedding representations of each candidate elderly care service is calculated, and the elderly care service recommendation results for the top preset number of target elderly persons are output according to the ranking of the matching degrees. This invention achieves effective denoising of the elderly care knowledge graph through a complete process of "knowledge fusion - dual denoising - multi-task optimization," thereby achieving accurate recommendation of elderly care services.

[0084] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0085] According to the present invention, a personalized recommendation method for elderly care services involves performing multiple rounds of noise reduction on an elderly person's preference knowledge graph to obtain a target elderly care knowledge graph, including: The first round of noise trimming was performed on the knowledge graph of elderly people's preferences for elderly care, and redundant edges were removed to obtain a preliminary denoised knowledge graph of elderly care. A second round of knowledge denoising is performed on the initially denoised elderly care knowledge graph. Based on the statistical independence criterion, redundant information that is irrelevant to the current care needs of the target elderly is filtered out to obtain the target elderly care knowledge graph.

[0086] Specifically, in some embodiments, step 203 of multi-round noise reduction includes the following steps: First round of noise reduction: The first round of noise reduction is performed on the knowledge graph of elderly people's preferences for elderly care, and redundant edges with a correlation degree lower than the dynamic threshold are removed to obtain a preliminary denoised knowledge graph of elderly care.

[0087] For example, the importance of each edge in the elderly care preference knowledge graph can be calculated in multiple dimensions, low-importance edges can be pruned according to dynamic thresholds, and the discard probability can be adjusted by random reparameterization to obtain a preliminary denoised elderly care knowledge graph.

[0088] The second round of denoising: A second round of knowledge denoising is performed on the initially denoised elderly care knowledge graph. Based on the statistical independence criterion, redundant information irrelevant to the current care needs of the target elderly is filtered out. Afterwards, elderly care standards verification is performed to obtain the target elderly care knowledge graph.

[0089] The statistical independence criterion is an information theory-based data filtering mechanism. It automatically identifies and filters redundant data that is mathematically independent of current care needs by calculating the mutual information or correlation coefficients between entities, relationships, and the real-time health status of the target elderly person in the knowledge graph. This ensures that the data input into the recommendation model is highly relevant and well-suited core care knowledge. In the statistical independence criterion, if two variables (e.g., the elderly person's "health status" and a certain "care knowledge") are statistically independent, it means that knowing information about one variable (e.g., the elderly person has diabetes) makes it impossible to infer information about the other variable (e.g., the parameters of a certain wheelchair).

[0090] The optimization objectives for the second round of knowledge denoising include: (1) Maximize relevance to care needs: The knowledge retained must be strongly related to the elderly’s current health status (e.g., “high risk of falling”) (e.g., “fall prevention care program”). (2) Minimize redundancy; the knowledge retained must be strongly correlated with the elderly person’s current health status (e.g., “high risk of falling”) (e.g., “fall prevention care program”). Prioritize retaining information necessary for elderly care.

[0091] In the noise reduction module, the principle followed by this invention is to retain only those knowledge that is not statistically independent of the "current care needs of the elderly" (i.e., there is a strong dependency relationship); and to remove those noise information that is statistically independent (i.e. irrelevant).

[0092] In practical applications, by optimizing the target filter to remove redundant elderly care knowledge, only information strongly related to care needs is retained, resulting in the final denoised elderly care knowledge graph (target elderly care knowledge graph). The specific operation method is as follows: The degree of dependence between each entity / relation in the initial denoised map and the "elderly profile vector" is measured through mathematical calculations (such as mutual information).

[0093] If the degree of dependence is high (violating independence), it indicates that the knowledge is necessary for the current care needs (e.g., elderly fracture -> associated with "orthopedic care").

[0094] If the degree of dependence is low (tending to be independent), it indicates that the knowledge is redundant (e.g., the elderly person has high blood pressure, but the atlas is associated with "the latest drug research on Alzheimer's disease").

[0095] It is important to emphasize that when calculating the statistical independence criterion, essential information for elderly care should be retained first, and key elderly care safety-related information should not be deleted in the name of noise reduction.

[0096] Optionally, if the final denoised elderly care knowledge graph contains association errors, in addition to automatically deleting the erroneous associations, the encoding scores of other elderly care entities associated with the erroneous edge need to be recalculated to avoid error propagation; when completing missing associations, the association scores are calculated using entity encoding functions to ensure the rationality of the association relationships.

[0097] Optionally, the elderly care information bottleneck denoising module can periodically trigger the re-verification of the final denoised elderly care knowledge graph in conjunction with the update frequency of elderly care standards, to ensure that the relationships such as "disease-care plan" and "assistive device-applicable population" in the graph are consistent with the latest standards, and avoid recommending outdated elderly care services.

[0098] The method provided in this invention employs a dual denoising module of "elderly care noise trimming + elderly care information bottleneck denoising (knowledge denoising)" to accurately filter redundant information in the elderly care knowledge graph. This addresses the problem of "interaction signals dominating and insufficient elderly care knowledge encoding" in traditional models, avoiding the recommendation of irrelevant care services to the elderly and significantly improving the utilization rate of elderly care knowledge in the embedding. In other words, knowledge denoising is accurate, and utilization is significantly improved.

[0099] According to a personalized recommendation method for elderly care services provided by the present invention, a first round of noise trimming is performed on the knowledge graph of elderly people's preferences for elderly care, removing redundant edges in the knowledge graph to obtain a preliminarily denoised knowledge graph of elderly care, including: The importance score of each edge in the knowledge graph of elderly preferences for elderly care is calculated. The edge importance score is obtained through multi-dimensional evaluation based on entity relevance, elderly health matching degree, and elderly care fit degree. Entity relevance is determined by the semantic distance between entities in the knowledge graph of elderly preferences for elderly care. Elderly health matching degree is determined by the correlation between the disease diagnosis in the target elderly’s electronic medical record and the elderly care entity. Elderly care fit degree is determined by the degree of fit between the elderly’s living habits, living environment and care entity. Based on the edge importance score corresponding to each edge, edges with an importance score lower than the dynamic threshold are identified as redundant edges and removed, resulting in a preliminary denoised elderly care knowledge graph.

[0100] Specifically, in some embodiments, the first round of noise reduction for the knowledge graph of elderly people's preferences for elderly care includes the following steps: First, calculate the edge importance score corresponding to each edge in the knowledge graph of elderly people's preferences for elderly care.

[0101] The edge importance calculation method includes a multi-dimensional assessment based on entity relevance, elderly health matching degree, and elderly care fit degree. Specifically, elderly care entity relevance is calculated through semantic distance between entities in the elderly care knowledge graph; elderly health matching degree is calculated through the correlation between disease diagnoses in the elderly's electronic medical records and elderly care entities; and elderly care fit degree is calculated through the degree of fit between the elderly's lifestyle habits, living environment, and care entities.

[0102] The dynamic threshold can be set to a specific quantile of edge importance in each round of training, pruning a certain proportion of low-importance edges, and the pruning process must avoid the associations necessary for elderly care, such as the association between disabled elderly and turning care programs.

[0103] Furthermore, based on the edge importance score corresponding to each edge, edges with an edge importance score lower than the dynamic threshold are identified as redundant edges (low-association edges) and removed, thereby obtaining a preliminary denoised elderly care knowledge graph and achieving preliminary denoising of the elderly preference elderly care knowledge graph.

[0104] The method provided in this invention calculates the importance of edges in an elderly care preference knowledge graph, and prunes redundant and low-association edges according to a dynamic threshold to obtain a preliminary denoised elderly care knowledge graph. Through a dual module of "elderly care noise pruning + elderly care information bottleneck denoising," redundant information in the elderly care knowledge graph is accurately filtered.

[0105] According to the present invention, a personalized recommendation method for elderly care services integrates the personal profile information of the target elderly person with an elderly care knowledge graph to construct an elderly care knowledge graph representing the personalized care needs of the target elderly person, including: Identify the personal profile information of the target elderly person and the multi-hop paths between entities in the elderly care knowledge graph; If any target entity can be associated with the target entity through a multi-hop path and the matching degree between the elderly node and the target entity meets the preset matching degree threshold, a direct association edge is established between the elderly node and the target entity; the elderly node is constructed based on the personal profile information of the target elderly person. Based on the directly related edges and the elderly care knowledge graph, a knowledge graph of elderly preferences for elderly care is constructed.

[0106] Specifically, in some embodiments, the knowledge fusion in step 202 includes the following steps: First, identify the multi-hop paths between elderly nodes (constructed based on the personal profile information of the target elderly) and entities in the elderly care knowledge graph. If any target entity can be associated through the multi-hop path and the matching degree between the elderly node and the target entity meets the preset conditions (such as matching degree higher than 80%), then establish a direct association edge between the elderly node and the target entity to directly associate the two.

[0107] The association rules for linking elderly people to potential related elderly care entities through multi-hop paths include: if an elderly person associates with an elderly care entity through a multi-hop path such as elderly-disease-treatment program or elderly-doctor-specialty field, and the matching degree between the entity and the elderly person's current health status reaches a high level, then a direct association between the elderly person and the entity is established; if an elderly person associates with an elderly care entity through a multi-hop path such as elderly-basic disease-care program or elderly-physical function-assistive device, and the matching degree between the entity and the elderly person's risk assessment results reaches a high level, then a direct association between the elderly person and the entity is established.

[0108] Furthermore, based on the directly related edges and the elderly care knowledge graph, an elderly preference elderly care knowledge graph can be constructed. Specifically, when integrating the elderly care knowledge graph, weighting is performed by interaction frequency × health relevance, and also by combining the urgency of care needs × profile matching degree to ensure that highly relevant and highly adaptable elderly care entities are retained first.

[0109] Optionally, for elderly people with duplicate medical records, the weight of the association edge between the elderly person and "disease-attending physician" and "disease-long-term follow-up project" is strengthened, and the weight adjustment is dynamically determined based on the entity coding score; for elderly people with long-term care needs, the weight of the association edge between them and "basic disease-dedicated caregiver" and "risk level-prevention and control care project" is strengthened.

[0110] The method provided in this embodiment of the invention, in the elderly care knowledge fusion module, associates elderly nodes with target entities according to association rules to generate a fused elderly preference elderly care knowledge graph, which can accurately match the personalized elderly care needs of different elderly people.

[0111] According to the present invention, a personalized recommendation method for elderly care services integrates the personal profile information of the target elderly person with an elderly care knowledge graph to construct an elderly care knowledge graph representing the personalized care needs of the target elderly person, including: Based on the individual profile information of the target elderly, structured information including risk assessment reports, vital sign monitoring lists, and initial care item lists is generated; By integrating structured information with an elderly care knowledge graph, a knowledge graph of elderly people's preferences for elderly care is obtained.

[0112] Specifically, in some embodiments, the generation of a knowledge graph of elderly people's preferences for elderly care includes a step of generating multi-dimensional data (structured information) on elderly care.

[0113] In practical applications, based on the target elderly person's personal profile information and the elderly care knowledge graph, three types of output data analysis risk assessment reports are generated: risk levels for falls and pressure sores, a recommended daily vital sign monitoring list specifying monitoring indicators and frequencies, and an appropriate initial daily care item list. Simultaneously, historical data entered by users from medical records and physical examination reports, or targeted physical examination suggestion templates generated by the system, are integrated into structured data and added to the knowledge graph fusion stage. That is, the structured data is associated and matched with entities in the elderly care knowledge graph, updating the elderly's preferred elderly care knowledge graph.

[0114] The method provided in this invention generates multi-dimensional data based on the personal profile information of the target elderly person and then supplements the multi-dimensional data into the elderly person's preference elderly care knowledge graph. This facilitates the subsequent precise recommendation of elderly care services based on the updated elderly person's preference elderly care knowledge graph, thereby improving the accuracy of the recommendation.

[0115] According to the present invention, a personalized recommendation method for elderly care services is provided, which, based on a target elderly care knowledge graph and elderly-care service interaction data, is jointly trained through a multi-task learning model to output the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, including: Based on the target elderly care knowledge graph and elderly-care service interaction data, a nursing interaction view and an elderly preference elderly care view are constructed. Graph attention networks are then trained on the nursing interaction view and the elderly preference elderly care view to obtain nursing interaction embedding representations and elderly care knowledge embedding representations, respectively. Nursing interaction embeddings and elderly care knowledge embeddings are mapped to the same semantic space and compared and aligned to enhance the similarity between positive sample pairs and reduce the distance between negative sample pairs, resulting in aligned knowledge embeddings. The definitions of positive and negative sample pairs conform to the constraints of the elderly care scenario. Positive sample pairs include at least one of the following: nursing interaction embeddings and elderly care knowledge embeddings for the same elderly person; nursing project embeddings and nursing staff expertise embeddings for the same disease; nursing needs embeddings and nursing project embeddings for the same elderly person; and rehabilitation nursing embeddings and nursing staff expertise embeddings for the same underlying disease. Negative sample pairs include at least one of the following: elderly person embeddings and non-health-related elderly care entity embeddings; elderly care entity embeddings and non-adaptive nursing entity embeddings for different diseases; and embeddings of elderly person embeddings and wheelchair-adaptive services for elderly people with mobility impairments. Based on the aligned knowledge embeddings corresponding to the target elderly care knowledge graph and the elderly-care service interaction data, and based on multi-task loss and elderly care compliance constraints, the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service are jointly trained through an adaptive momentum optimizer.

[0116] Specifically, in some embodiments, the multi-task optimization process includes the following steps: First, based on the target elderly care knowledge graph and elderly-care service interaction data, construct a care interaction view and an elderly preference elderly care view; for example, randomly discard a certain proportion of non-critical interaction edges in the elderly-elderly care service interaction graph; generate the elderly preference elderly care view: retain the core elderly care relationships in the preliminarily denoised elderly care knowledge graph.

[0117] Furthermore, graph attention networks are trained on the nursing interaction view and the elderly preference elderly care view to obtain nursing interaction embedding representations and elderly care knowledge embedding representations. Specifically, two-layer graph attention networks are used to train the two views respectively. The attention score calculation is combined with the weight of elderly care entities. For the class edges of "elderly-caregiver" and "elderly-care project", the attention score is additionally multiplied by the credibility of the elderly care entity to output the nursing interaction embedding and the elderly care knowledge embedding.

[0118] Optionally, the training of the dual-view elderly care graph attention network can incorporate "elderly care time features": for elderly-elderly care service interaction edges in the recent period, the attention score is multiplied by a specific coefficient to prioritize capturing recent care preferences; for expired medicines, out-of-service caregivers, and other invalid entities, their embedding weights are automatically marked and reduced to avoid invalid recommendations.

[0119] Furthermore, the nursing interaction embedding representation and the elderly care knowledge embedding representation are mapped to the same semantic space and compared and aligned. This strengthens the similarity between positive sample pairs and widens the distance between negative sample pairs, resulting in aligned knowledge embeddings.

[0120] The definitions of positive and negative sample pairs conform to the constraints of the elderly care scenario. Positive sample pairs include at least one of the following: the embedding of nursing interaction with the same elderly person and the embedding of elderly care knowledge; the embedding of nursing projects related to the same disease and the embedding of nursing staff expertise; the embedding of nursing needs with the same elderly person and the embedding of nursing projects; and the embedding of rehabilitation nursing related to the same underlying disease and the embedding of nursing staff expertise. Negative sample pairs include at least one of the following: the embedding of the elderly person with the embedding of non-health-related elderly care entities; the embedding of elderly care entities related to different diseases; the embedding of the elderly person with the embedding of non-fit nursing entities; and the embedding of elderly people with mobility and wheelchair-fitting services.

[0121] The similarity of sample pairs can be calculated using scaled cosine similarity, with additional verification of elderly care standards added to the similarity calculation of high-risk elderly care entities. Embedding alignment can be optimized through contrastive loss, maximizing the similarity of positive samples and minimizing the distance between negative samples to align elderly behavioral signals with elderly care knowledge signals. A nursing stage adaptation mechanism is added: for elderly individuals in the acute phase of illness, the learning weights of emergency service embedding and rapid examination item embedding are increased by a certain proportion, with weight adjustments dynamically allocated through the loss weights in the total loss formula; for elderly individuals in the recovery phase of illness, the learning weights of rehabilitation therapy embedding and family doctor embedding are strengthened; for elderly individuals in the stable phase of chronic disease, the learning accuracy of chronic disease management embedding and regular follow-up item embedding is optimized; for elderly individuals in the postoperative recovery phase, the weights of postoperative care embedding and rehabilitation training embedding are increased; for frail elderly individuals, the learning weights of nutritional support embedding and complication prevention care embedding are strengthened; the switching between different nursing stages is automatically triggered based on the health status score in the elderly individual's latest examination report, and the corresponding encoded score of the elderly care entity is recalculated.

[0122] Furthermore, based on the aligned knowledge embeddings corresponding to the target elderly care knowledge graph and the elderly-care service interaction data, and using a multi-task loss algorithm through joint training with an adaptive momentum optimizer, the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service are output.

[0123] The aligned knowledge embeddings correspond to the target elderly care knowledge graph, which is an updated elderly care knowledge graph based on the aligned knowledge embeddings. During training, compliance constraints for elderly care can be incorporated. These constraints include: if a recommended nursing service requires specific qualifications, such as rehabilitation therapy, the embedding learning must be associated with the nursing staff's qualification characteristics; if a recommended assistive device requires doctor evaluation, it must be associated with the elderly person's doctor evaluation record characteristics; if no evaluation record exists, the embedding weight of the device is reduced by a certain percentage; if a recommended treatment requires specific qualifications, the embedding learning must be associated with the doctor's qualification characteristics; if a recommended medication is a prescription drug, it must be associated with the elderly person's prescription record characteristics; if no prescription record exists, the embedding weight of the drug is reduced by a certain percentage.

[0124] In practical applications, a total loss function is constructed by integrating multi-task losses, and an adaptive momentum optimizer is used for joint training. The training is iteratively continued until the total loss converges, outputting the final embedding of the elderly and elderly care services. The loss weights of the multi-task training can be dynamically adjusted according to different elderly care scenarios to ensure the model adapts to the needs of each scenario. During training, the effectiveness of compliance constraints in elderly care can be periodically verified, and if compliance deviations are found, the weights of the associated features in the embedding learning are adjusted promptly.

[0125] The method provided in this invention supplements the features of long-tail elderly care services by embedding elderly care knowledge, thereby improving the coverage of long-tail service recommendations. For cold-start entities such as newly hired caregivers and newly added care technologies, effective embeddings are quickly generated based on knowledge graph associations, improving the accuracy of cold-start entity recommendations and balancing the allocation of elderly care resources.

[0126] According to the present invention, a personalized recommendation method for elderly care services includes a multi-task loss comprising personalized recommendation loss for elderly care, comparison and alignment loss for elderly care, and denoising loss for elderly care bottlenecks.

[0127] Specifically, by integrating multi-task losses (including personalized elderly care recommendation loss, elderly care comparison alignment loss, and elderly care bottleneck denoising loss), an adaptive momentum optimizer is used for joint training. The training is iteratively completed until the total loss converges, and the final embedding of the elderly and elderly care services is output.

[0128] The method provided in this invention rapidly generates effective embeddings based on knowledge graph associations, thereby improving recommendation accuracy.

[0129] According to a personalized recommendation method for elderly care services provided by the present invention, the method outputs the recommended results of elderly care services for the first preset number of target elderly people based on the ranking of each matching degree, including: The system outputs the recommended elderly care services for the first preset number of target elderly individuals, along with corresponding personalized health advice and traceable explanations of the recommended services, in descending order of matching degree. The recommended elderly care services for the first preset number of target elderly individuals comply with elderly care standards.

[0130] Specifically, in some embodiments, step 205 is implemented through the following steps: Output the top-K recommended elderly care services in descending order of matching degree, along with corresponding personalized health advice and traceable explanations of the elderly care recommendations.

[0131] Personalized health recommendations are generated based on the final embedded matching results (various matching degrees) between the elderly and care services, combined with risk assessment reports and daily vital sign monitoring lists, and include diet, exercise, medication reminders, and nursing operation standards.

[0132] The recommended elderly care services include the following: the matching points between the recommended elderly care services and the elderly person's health condition; the applicable scope and contraindications of the elderly care services; if the recommended service is an invasive procedure, additional alternatives will be provided; if the recommended medication is a prescription drug, relevant precautions and dosage references will be indicated; the operational difficulty of the recommended elderly care services and the suitable scenarios (home / community); if the recommended service requires family cooperation, additional points on family cooperation will be provided; all descriptions are linked to the corresponding elderly care knowledge graph entities to ensure the traceability of the recommendation basis.

[0133] In addition, a manual review process can be added before the output of elderly care service recommendations to conduct a second verification of high-risk elderly care services, ensuring that the recommendations comply with elderly care standards and further reducing recommendation risks. The Top-K value can be dynamically adjusted based on recommendation performance feedback to ensure recommendation accuracy while taking into account the diversity of elderly care services.

[0134] The method provided in this invention outputs multiple elderly care service recommendations for the target elderly person, along with corresponding personalized health advice and traceable elderly care recommendation descriptions, ordered in descending order of matching degree. The recommendation descriptions are linked to entities in the elderly care knowledge graph, elderly care standards, and the qualifications of caregivers. It supports log pruning queries and QR code traceability, solving the "black box" problem of traditional recommendations, meeting the verification needs of quality control and regulatory departments in elderly care institutions, and improving the transparency of elderly care services and the trust of the elderly and their families.

[0135] According to a personalized recommendation method for elderly care services provided by the present invention, the elderly care knowledge graph is updated through the following steps, including: The knowledge graph of elderly care is updated based on regularly synchronized external data sources. The semantic distances between entities in the updated elderly care knowledge graph are recalculated, and the recalculated semantic distances between entities are updated in the metadata storage layer of the updated elderly care knowledge graph.

[0136] Specifically, in some embodiments, the elderly care knowledge graph updating process includes the following steps: In practical applications, the elderly care knowledge graph is first updated based on regularly synchronized external data sources, such as the latest drug catalogs, clinical guidelines, elderly care standards, and assistive device standards. After updating the knowledge graph, the semantic distances between entities are recalculated, and these recalculated semantic distances are then updated in the metadata storage layer of the updated knowledge graph to ensure its timeliness.

[0137] Optionally, for elderly people with duplicate medical records, the weight of the association edges between the elderly person and their attending physician and long-term follow-up project is strengthened, and the weight adjustment is dynamically determined based on the entity coding score; for elderly people with long-term care needs, the weight of the association edges between the elderly person and their dedicated caregiver for basic diseases and risk level prevention and control care project is strengthened.

[0138] The method provided in this invention regularly updates the elderly care knowledge graph, ensuring that the elderly care knowledge graph used for personalized elderly care service recommendations is the latest one, thereby improving the accuracy of recommending personalized elderly care services to target elderly individuals.

[0139] Optionally, the target elderly care knowledge graph can be validated against elderly care standards: if there is an incorrect association between disease and contraindicated drugs in the graph, the associated edge is automatically deleted and the source of the error is recorded, and the coding scores of other elderly care entities associated with the incorrect edge are recalculated; if there is a missing association between treatment item and required qualification, the associated edge is automatically completed and the association score is calculated using the entity coding function; if there is an incorrect association between elderly risk level and contraindicated care items in the graph, the edge is automatically deleted and the source is recorded; if there is a missing association between special care item and required care qualification, the edge is automatically completed and the association score is calculated, ensuring that the graph complies with clinical treatment standards and elderly care standards.

[0140] Optionally, during the training of the graph attention network, the geographical features of elderly care entities can be associated with it: For elderly people in areas with a high incidence of endemic diseases, their attention scores on the edges related to endemic disease specialists and endemic disease treatment programs are multiplied by a specific coefficient; for elderly people seeking medical treatment across regions, the weights on the edges related to designated hospitals under medical insurance in other regions and referral channels are strengthened, and the weight coefficients are set to specific values. For elderly people living alone, their attention scores related to home care services and emergency call services are multiplied by a specific coefficient; for elderly people receiving community care, the weight of their associations with community care centers and neighborhood mutual assistance services is strengthened. During the training process, the list of regional medical centers released by relevant departments is synchronized in real time. For elderly care entities associated with regional medical centers, their attention scores during the embedding learning are multiplied by a specific coefficient to ensure that high-level referral resources are embedded and learned first.

[0141] Optionally, a nursing phase adaptation mechanism can be incorporated into multi-task training: For elderly patients in the acute phase of illness, the learning weights for embedding emergency services and rapid examination items are increased by a certain proportion, and the weight adjustment is achieved through dynamic allocation of loss weights in the total loss formula; for elderly patients in the recovery phase of illness, the learning weights for embedding rehabilitation therapy and family doctor are strengthened; for elderly patients in the stable phase of chronic disease, the learning accuracy of embedding chronic disease management and regular follow-up items is optimized; for elderly patients in the postoperative recovery phase, the weights for embedding postoperative care and rehabilitation training are increased; for frail elderly patients, the learning weights for embedding nutritional support and complication prevention care are strengthened. The switching between different care stages is automatically triggered based on the health status score in the elderly person's latest examination report, and the corresponding coded score of the elderly care entity is recalculated.

[0142] Optionally, the edge discarding decision in the initial denoising adopts a stochastic reparameterization method: achieving differentiable computation, and adjusting the discarding probability through activation function and temperature parameter, wherein the temperature parameter is dynamically adjusted according to the risk level of the elderly care entity: a smaller temperature parameter is used for the edges related to high-risk elderly care entities, and a larger temperature parameter is used for the edges related to regular elderly care entities, to ensure the stability of the decision of risky elderly care associations.

[0143] For example, Figure 3 This is the second flowchart illustrating the personalized recommendation method for elderly care services provided by this invention, as shown below. Figure 3 As shown, the method includes: Step 301: Input the elderly person's personal profile data; Step 302: Generate multidimensional data based on the profile data and the pre-built elderly care knowledge graph; Step 303: Integrate multi-source data to generate a preference map and strengthen the association with specific elderly individuals; Step 304: Calculate edge importance in multiple dimensions and prune the edges, and adjust the discard probability to obtain a preliminary denoised graph; Step 305: Dual-view embedding and mapping alignment, plus a nursing stage adaptation mechanism to dynamically adjust weights; Step 306: Filter redundant knowledge and perform compliance verification to obtain the final denoised map; Step 307: Integrate loss joint training with compliance constraints to output the final embedded data of elderly care services; Step 308: Calculate the matching degree and output the recommendation results. High-risk services are manually reviewed and can be retrospectively re-recommended.

[0144] The personalized elderly care service recommendation device provided by the present invention is described below. The personalized elderly care service recommendation device described below and the personalized elderly care service recommendation method described above can be referred to in correspondence.

[0145] Figure 4This is a schematic diagram of the personalized elderly care service recommendation device provided by the present invention, as shown below. Figure 4 As shown, the personalized elderly care service recommendation device 400 includes: The acquisition unit 410 is used to acquire the personal profile information of the target elderly person; the personal profile information includes at least one of basic information, health data, and care preferences. The knowledge fusion unit 420 is used to fuse the personal profile information of the target elderly with the elderly care knowledge graph to construct an elderly preference elderly care knowledge graph that represents the personalized care needs of the target elderly. The multi-round denoising unit 430 is used to perform multi-round denoising on the elderly preference elderly care knowledge graph to obtain the target elderly care knowledge graph. The multi-task optimization unit 440 is used to jointly train a multi-task learning model based on the knowledge graph of the target elderly care and the elderly-care service interaction data, and output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. The personalized recommendation unit 450 is used to calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

[0146] The apparatus provided in this embodiment of the invention includes an acquisition unit 410 for acquiring personal profile information of a target elderly person, wherein the personal profile information includes at least one of basic information, health data, and care preferences; a knowledge fusion unit 420 for fusing the personal profile information of the target elderly person with an elderly care knowledge graph to construct an elderly person preference elderly care knowledge graph representing the personalized care needs of the target elderly person; further, a multi-round denoising unit 430 for performing multi-round denoising on the elderly person preference elderly care knowledge graph to obtain a target elderly care knowledge graph; a multi-task optimization unit 440 for jointly training a multi-task learning model based on the target elderly care knowledge graph and elderly-care service interaction data, and outputting the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service; and finally, a personalized recommendation unit 450 for calculating the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and outputting the elderly care service recommendation results for the first preset number of target elderly persons according to the ranking of the matching degrees. This invention achieves effective noise reduction of the elderly care knowledge graph through a full process of "knowledge fusion - dual denoising - multi-task optimization", thereby enabling accurate recommendation of elderly care services.

[0147] According to the present invention, a personalized recommendation device 400 for elderly care services includes a multi-round noise reduction unit 430, which is specifically used for: The first round of noise trimming is performed on the elderly preference elderly care knowledge graph to remove redundant edges, resulting in a preliminary denoised elderly care knowledge graph. A second round of knowledge denoising is performed on the initially denoised elderly care knowledge graph. Based on the statistical independence criterion, redundant information that is irrelevant to the current care needs of the target elderly is filtered out to obtain the target elderly care knowledge graph.

[0148] According to the present invention, a personalized recommendation device 400 for elderly care services is provided, wherein the multi-round noise reduction unit 430 is further used for: The importance score of each edge in the knowledge graph of elderly preferences for elderly care is calculated. The importance score is obtained through multi-dimensional evaluation based on entity relevance, elderly health matching degree, and elderly care fit degree. The entity relevance is determined by the semantic distance between entities in the knowledge graph of elderly preferences for elderly care. The elderly health matching degree is determined by the correlation between the disease diagnosis in the target elderly’s electronic medical record and the elderly care entity. The elderly care fit degree is determined by the degree of fit between the elderly’s living habits, living environment and care entity. Based on the edge importance score corresponding to each edge, edges with an edge importance score lower than the dynamic threshold are identified as redundant edges and removed to obtain the preliminary denoised elderly care knowledge graph.

[0149] According to the present invention, a personalized recommendation device 400 for elderly care services is provided, wherein the knowledge fusion unit 420 is specifically used for: Identify the personal profile information of the target elderly person and the multi-hop paths between entities in the elderly care knowledge graph; If any target entity among the entities can be associated through the multi-hop path and the matching degree between the elderly node and the target entity meets the preset matching degree threshold, a direct association edge is established between the elderly node and the target entity; the elderly node is constructed based on the personal profile information of the target elderly. Based on the directly related edges and the elderly care knowledge graph, the elderly preference elderly care knowledge graph is constructed.

[0150] According to the present invention, a personalized recommendation device 400 for elderly care services is provided, wherein the knowledge fusion unit 420 is further used for: Based on the personal profile information of the target elderly, structured information including a risk assessment report, a vital sign monitoring list, and an initial care item list is generated; The structured information is fused with the elderly care knowledge graph to obtain the elderly preference elderly care knowledge graph.

[0151] According to the present invention, a personalized recommendation device 400 for elderly care services is provided, wherein the multi-task optimization unit 440 is specifically used for: Based on the target elderly care knowledge graph and the elderly-care service interaction data, a nursing interaction view and an elderly preference elderly care view are constructed. Graph attention networks are then trained on the nursing interaction view and the elderly preference elderly care view to obtain nursing interaction embedding representations and elderly care knowledge embedding representations, respectively. The nursing interaction embedding representation and the elderly care knowledge embedding representation are mapped to the same semantic space and compared and aligned to enhance the similarity between positive sample pairs and reduce the distance between negative sample pairs, resulting in aligned knowledge embeddings. The definitions of the positive and negative sample pairs conform to the constraints of the elderly care scenario. Positive sample pairs include at least one of the following: nursing interaction embedding and elderly care knowledge embedding for the same elderly person; nursing project embedding and nursing staff expertise embedding for the same disease; nursing needs embedding and nursing project embedding for the same elderly person; rehabilitation nursing embedding and nursing staff expertise embedding for the same underlying disease. Negative sample pairs include at least one of the following: elderly person embedding and non-health-related elderly care entity embedding; elderly care entity embedding and non-fitted nursing entity embedding for different diseases; elderly person embedding and wheelchair-fitted service embedding for elderly people with mobility impairments. Based on the aligned knowledge embeddings corresponding to the target elderly care knowledge graph and the elderly-care service interaction data, and using multi-task loss and elderly care compliance constraints, the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service are output through joint training with an adaptive momentum optimizer.

[0152] According to the present invention, a personalized recommendation device 400 for elderly care services includes a multi-task loss comprising personalized recommendation loss for elderly care, comparison and alignment loss for elderly care, and denoising loss for elderly care bottlenecks.

[0153] According to the present invention, a personalized recommendation device 400 for elderly care services is provided, wherein the personalized recommendation unit 450 is specifically used for: The system outputs multiple elderly care service recommendations, corresponding personalized health advice, and traceable elderly care recommendation descriptions, ordered in descending order of matching degree. The multiple elderly care service recommendations for the target elderly person comply with elderly care standards.

[0154] According to the present invention, a personalized recommendation device 400 for elderly care services updates the elderly care knowledge graph through the following steps: The elderly care knowledge graph is updated based on regularly synchronized external data sources. The semantic distances between entities in the updated elderly care knowledge graph are recalculated, and the recalculated semantic distances between entities are updated in the metadata storage layer of the updated elderly care knowledge graph.

[0155] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a personalized recommendation method for elderly care services, which includes: Obtain personal profile information of the target elderly person; the personal profile information includes at least one of the following: basic information, health data, and care preferences. By integrating the personal profile information of the target elderly with the elderly care knowledge graph, an elderly preference elderly care knowledge graph representing the personalized care needs of the target elderly is constructed. The elderly care preference knowledge graph is subjected to multiple rounds of noise reduction to obtain the target elderly care knowledge graph. Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

[0156] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0157] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the personalized recommendation method for elderly care services provided by the above methods, the method comprising: Obtain personal profile information of the target elderly person; the personal profile information includes at least one of the following: basic information, health data, and care preferences. By integrating the personal profile information of the target elderly with the elderly care knowledge graph, an elderly preference elderly care knowledge graph representing the personalized care needs of the target elderly is constructed. The elderly care preference knowledge graph is subjected to multiple rounds of noise reduction to obtain the target elderly care knowledge graph. Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

[0158] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a personalized recommendation method for elderly care services provided by the methods described above, the method comprising: Obtain personal profile information of the target elderly person; the personal profile information includes at least one of the following: basic information, health data, and care preferences. By integrating the personal profile information of the target elderly with the elderly care knowledge graph, an elderly preference elderly care knowledge graph representing the personalized care needs of the target elderly is constructed. The elderly care preference knowledge graph is subjected to multiple rounds of noise reduction to obtain the target elderly care knowledge graph. Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service; the elderly-care service interaction data is used to reflect the correspondence between different elderly people and elderly care services. Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

[0159] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0160] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for personalized recommendation of elderly care services, characterized in that, include: Obtain personal profile information of the target elderly person; The personal profile information includes at least one of the following: basic information, health data, and care preferences; By integrating the personal profile information of the target elderly with the elderly care knowledge graph, an elderly preference elderly care knowledge graph representing the personalized care needs of the target elderly is constructed. The elderly care preference knowledge graph is subjected to multiple rounds of noise reduction to obtain the target elderly care knowledge graph. Based on the target elderly care knowledge graph and elderly-care service interaction data, a multi-task learning model is jointly trained to output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service. The elderly-nursing service interaction data is used to reflect the correspondence between different elderly people and elderly care services; Calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

2. The personalized recommendation method for elderly care services according to claim 1, characterized in that, The process of performing multiple rounds of noise reduction on the elderly preference elderly care knowledge graph to obtain the target elderly care knowledge graph includes: The first round of noise trimming is performed on the elderly preference elderly care knowledge graph to remove redundant edges, resulting in a preliminary denoised elderly care knowledge graph. A second round of knowledge denoising is performed on the initially denoised elderly care knowledge graph. Based on the statistical independence criterion, redundant information that is irrelevant to the current care needs of the target elderly is filtered out to obtain the target elderly care knowledge graph.

3. The personalized recommendation method for elderly care services according to claim 2, characterized in that, The first round of noise trimming on the elderly preference elderly care knowledge graph involves removing redundant edges to obtain a preliminarily denoised elderly care knowledge graph, including: The importance score of each edge in the knowledge graph of elderly preferences for elderly care is calculated. The importance score is obtained through multi-dimensional evaluation based on entity relevance, elderly health matching degree, and elderly care fit degree. The entity relevance is determined by the semantic distance between entities in the knowledge graph of elderly preferences for elderly care. The elderly health matching degree is determined by the correlation between disease diagnosis and elderly care entity in the electronic medical record of the target elderly person. The elderly care fit degree is determined by the degree of fit between the elderly person's living habits, living environment and care entity. Based on the edge importance score corresponding to each edge, edges with an edge importance score lower than the dynamic threshold are identified as redundant edges and removed to obtain the preliminary denoised elderly care knowledge graph.

4. The personalized recommendation method for elderly care services according to any one of claims 1-3, characterized in that, The process of integrating the target elderly person's personal profile information with an elderly care knowledge graph to construct an elderly care preference knowledge graph representing the target elderly person's personalized care needs includes: Identify the personal profile information of the target elderly person and the multi-hop paths between entities in the elderly care knowledge graph; If any target entity among the entities can be associated through the multi-hop path and the matching degree between the elderly node and the target entity meets the preset matching degree threshold, a direct association edge is established between the elderly node and the target entity; the elderly node is constructed based on the personal profile information of the target elderly. Based on the directly related edges and the elderly care knowledge graph, the elderly preference elderly care knowledge graph is constructed.

5. The personalized recommendation method for elderly care services according to any one of claims 1-3, characterized in that, The process of integrating the target elderly person's personal profile information with an elderly care knowledge graph to construct an elderly care preference knowledge graph representing the target elderly person's personalized care needs includes: Based on the personal profile information of the target elderly, structured information including a risk assessment report, a vital sign monitoring list, and an initial care item list is generated; The structured information is fused with the elderly care knowledge graph to obtain the elderly preference elderly care knowledge graph.

6. The personalized recommendation method for elderly care services according to any one of claims 1-3, characterized in that, The process, based on the target elderly care knowledge graph and elderly-care service interaction data, involves joint training using a multi-task learning model to output the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, including: Based on the target elderly care knowledge graph and the elderly-care service interaction data, a nursing interaction view and an elderly preference elderly care view are constructed. Graph attention networks are then trained on the nursing interaction view and the elderly preference elderly care view to obtain nursing interaction embedding representations and elderly care knowledge embedding representations, respectively. The nursing interaction embedding representation and the elderly care knowledge embedding representation are mapped to the same semantic space and compared and aligned to enhance the similarity between positive sample pairs and reduce the distance between negative sample pairs, resulting in aligned knowledge embeddings. The definitions of the positive and negative sample pairs conform to the constraints of the elderly care scenario. Positive sample pairs include at least one of the following: nursing interaction embedding and elderly care knowledge embedding for the same elderly person; nursing project embedding and nursing staff expertise embedding for the same disease; nursing needs embedding and nursing project embedding for the same elderly person; rehabilitation nursing embedding and nursing staff expertise embedding for the same underlying disease. Negative sample pairs include at least one of the following: elderly person embedding and non-health-related elderly care entity embedding; elderly care entity embedding and non-fitted nursing entity embedding for different diseases; elderly person embedding and wheelchair-fitted service embedding for elderly people with mobility impairments. Based on the aligned knowledge embeddings corresponding to the target elderly care knowledge graph and the elderly-care service interaction data, and using multi-task loss and elderly care compliance constraints, the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service are output through joint training with an adaptive momentum optimizer.

7. The personalized recommendation method for elderly care services according to claim 6, characterized in that, The multi-task loss includes personalized elderly care recommendation loss, elderly care comparison alignment loss, and elderly care bottleneck denoising loss.

8. The personalized recommendation method for elderly care services according to claim 6, characterized in that, The step of sorting according to the matching degree and outputting multiple elderly care service recommendations corresponding to the target elderly person includes: The system outputs multiple elderly care service recommendations, corresponding personalized health advice, and traceable elderly care recommendation descriptions, ordered in descending order of matching degree. The multiple elderly care service recommendations for the target elderly person comply with elderly care standards.

9. The personalized recommendation method for elderly care services according to any one of claims 1-3, characterized in that, The elderly care knowledge graph is updated through the following steps: The elderly care knowledge graph is updated based on regularly synchronized external data sources; The semantic distances between entities in the updated elderly care knowledge graph are recalculated, and the recalculated semantic distances between entities are updated in the metadata storage layer of the updated elderly care knowledge graph.

10. A personalized recommendation device for elderly care services, characterized in that, include: The acquisition unit is used to acquire the personal profile information of the target elderly person. The personal profile information includes at least one of the following: basic information, health data, and care preferences; The knowledge fusion unit is used to fuse the personal profile information of the target elderly with the elderly care knowledge graph to construct an elderly preference elderly care knowledge graph that represents the personalized care needs of the target elderly. A multi-round denoising unit is used to perform multi-round denoising on the elderly preference elderly care knowledge graph to obtain the target elderly care knowledge graph. The multi-task optimization unit is used to jointly train the target elderly care knowledge graph and elderly-care service interaction data through a multi-task learning model, and output the final embedding representation of the target elderly and the final embedding representation of each candidate elderly care service. The elderly-nursing service interaction data is used to reflect the correspondence between different elderly people and elderly care services; A personalized recommendation unit is used to calculate the matching degree between the final embedding representation of the target elderly person and the final embedding representation of each candidate elderly care service, and output multiple elderly care service recommendation results corresponding to the target elderly person according to the ranking of the matching degrees.

Images