A method and system for intelligent matching of health care services based on multi-source big data

By decomposing and describing vector extraction of multi-source data sets, combining portrait label inference and impact coefficient adjustment, the problem of insufficient classification accuracy caused by the complexity of multi-source data in the prior art is solved, and the matching accuracy of health care services is improved.

CN118690087BActive Publication Date: 2024-11-01贵阳康养职业大学 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411170457.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-01
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

The existing intelligent matching methods for health care services based on multi-source big data lead to insufficient classification accuracy in data complexity, which in turn affects the accuracy of service matching.

Method used

By decomposing the multi-source data set of the target user, obtaining multiple data subsets, and performing description vector extraction and portrait label inference for each data subset, determining the influence coefficient, adjusting the subset description vector, and finally performing the first Secondary portrait label reasoning to improve the accuracy of user portraits.

Benefits of technology

The matching accuracy of health care services is improved, ensuring that health care services are more accurately matched based on user portrait reasoning results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118690087B_ABST
    Figure CN118690087B_ABST
Patent Text Reader

Abstract

The present invention provides a method and system for intelligent matching of health care services based on multi-source big data, decomposes the target data to obtain multiple data subsets, extracts description vectors, obtains subset description vectors, and performs first portrait label reasoning on each data subset according to each subset description vector to obtain a data subset portrait label; thereby determining each influence coefficient according to each data subset portrait label, adjusting each subset description vector to obtain an adjusted subset description vector, and finally performing second portrait label reasoning on the target data set according to each adjusted subset description vector, obtaining a user portrait reasoning result of the target data set, and matching health care services. Because the influence coefficient of adjusting each subset description vector represents the importance of the data subset to the target data set, each adjusted subset description vector obtained by adjustment can more accurately characterize the target data set, and the user portrait reasoning result is more accurate, thereby improving the reasonable matching of health care services.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the fields of data processing and machine learning, and more specifically, to a method and system for intelligent matching of health care services based on multi-source big data. Background Art

[0002] With the continuous improvement of people's health awareness and the intensification of population aging, the demand for health care services is gradually increasing. Traditional health care services mainly rely on manual matching, which has problems such as low efficiency and inaccurate matching. In order to solve these problems, intelligent matching methods for health care services based on multi-source big data have emerged. Multi-source big data includes users' health records, medical records, social network data, etc. By analyzing and mining these data, users' health status, living habits, interests and hobbies can be obtained, so as to achieve more accurate health care service matching. At present, some intelligent matching methods for health care services based on multi-source big data have been proposed. These methods usually use machine learning, data mining and other technologies to analyze and process multi-source big data to achieve intelligent matching. However, these methods still have some problems. For example, due to the complexity and diversity of the data, the overall classification accuracy is insufficient, which in turn causes the problem of low overall service matching accuracy. Summary of the invention

[0003] In view of this, the embodiments of the present application at least provide a method and system for intelligent matching of health care services based on multi-source big data.

[0004] According to one aspect of an embodiment of the present disclosure, a method for intelligent matching of health care services based on multi-source big data is provided, the method comprising:

[0005] Decomposing a target user multi-source data set to obtain a plurality of data subsets included in the target user multi-source data set;

[0006] Extracting description vectors for each of the data subsets to obtain subset description vectors for each of the data subsets;

[0007] According to the subset description vector of each of the data subsets, first image label inference is performed on each of the data subsets to obtain a data subset image label of each of the data subsets;

[0008] For each of the data subsets, determining an influence coefficient of the data subset according to the data subset portrait label of the data subset, and determining an adjusted subset description vector of the data subset according to the subset description vector of the data subset and the influence coefficient, wherein the influence coefficient is used to indicate the importance of the data subset to the user portrait inference result of the target user multi-source data set;

[0009] Performing a second portrait label inference on the target user multi-source data set according to the adjusted subset description vectors of the multiple data subsets to obtain a user portrait inference result of the target user multi-source data set;

[0010] Based on the user portrait inference result, matching health care services are determined in a preset health care service list.

[0011] According to an example of an embodiment of the present disclosure, the extracting description vectors of each of the data subsets to obtain the subset description vectors of each of the data subsets includes:

[0012] Extracting the first-level description vector of each of the data subsets to obtain the first-level hidden subset description vector of each of the data subsets;

[0013] Merging the first-level hidden subset description vectors of the respective multiple data subsets to obtain an array of hidden subset description vectors of the target user multi-source data set;

[0014] A second-level description vector extraction is performed on the hidden subset description vector array to obtain a subset description vector array of the target user multi-source data set, wherein the subset description vector array includes the second-level subset description vectors of each of the data subsets, and the dimension of the second level is smaller than the dimension of the first level.

[0015] According to an example of an embodiment of the present disclosure, determining the influence coefficient of the data subset based on the data subset portrait label of the data subset includes:

[0016] Obtaining the number of label types inferred by the first portrait label, and determining an equiprobability distribution of the number of label types;

[0017] Determining an error between a data subset image label of the data subset and the equal probability distribution;

[0018] determining the error as an influence coefficient of the data subset;

[0019] The step of determining an adjusted subset description vector of the data subset based on the subset description vector of the data subset and the influence coefficient includes:

[0020] Multiplying the subset description vector of the data subset by the influence coefficient of the data subset to obtain the adjusted subset description vector of the data subset;

[0021] The performing second portrait label inference on the target user multi-source data set based on the adjusted subset description vectors of the multiple data subsets to obtain a user portrait inference result of the target user multi-source data set includes:

[0022] Performing a description vector fusion operation on the adjusted subset description vectors of the multiple data subsets to obtain a fused subset description vector;

[0023] According to the fused subset description vector, a second portrait label inference is performed on the target user multi-source data set to obtain a user portrait inference result of the target user multi-source data set.

[0024] According to an example of an embodiment of the present disclosure, the first portrait label reasoning is performed based on a data point level mapping reasoning component of a data set portrait label network, and the second portrait label reasoning is performed based on a data set level mapping reasoning component of the data set portrait label network;

[0025] The method further comprises:

[0026] Acquire a user multi-source training data set combination for debugging the data set portrait label network, wherein the user multi-source training data set combination includes multiple user multi-source training data sets, the user multi-source training data sets carry data set annotation information, and the user multi-source training data sets include multiple training data subsets, and the training data subsets carry data subset annotation information;

[0027] Based on the data point level mapping reasoning component, first portrait label reasoning is performed on each of the training data subsets in each of the user multi-source training data sets to obtain data point reasoning results for each of the training data subsets;

[0028] For each of the user multi-source training data sets, the following operations are performed:

[0029] In combination with the data subset annotation information of each of the training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set;

[0030] Determine the network loss of the data set portrait label network according to the errors between the data point inference results of each of the training data subsets and the data set annotation information, and the errors between the data set inference results of the user multi-source training data set and the data set annotation information;

[0031] The network internal configuration variables of the data set portrait label network are adjusted according to the network loss to debug the data set portrait label network.

[0032] According to an example of an embodiment of the present disclosure, the determining of the network loss of the data set portrait label network based on the data point inference results of each of the training data subsets and the errors of the data set annotation information, and the data set inference results of the user multi-source training data set and the errors of the data set annotation information, includes:

[0033] For each of the training data subsets, determining the hidden data point level loss of the data set portrait labeling network according to the data point inference result of the training data subset and the error of the data subset annotation information;

[0034] Sum the hidden data point level losses of the data set portrait label network to obtain the data point level loss of the data set portrait label network;

[0035] Determine a dataset level loss of the dataset portrait label network according to the dataset inference result of the user multi-source training dataset and the error of the dataset annotation information;

[0036] According to the first loss influence coefficient of the data point level loss and the second loss influence coefficient of the data set level loss, the data point level loss and the data set level loss are fused to obtain the network loss of the data set portrait label network;

[0037] The data point level mapping reasoning component performs first portrait label reasoning on each of the training data subsets in each of the user multi-source training data sets to obtain data point reasoning results for each of the training data subsets, including:

[0038] Obtaining a training data subset description vector of each of the training data subsets in each of the user multi-source training data sets;

[0039] Based on the data point level mapping reasoning component, according to the training data subset description vector of each training data subset, first image label reasoning is performed on each training data subset to obtain data point reasoning results of each training data subset;

[0040] The combining the data subset annotation information of each of the training data subsets in the user multi-source training data set, performing second portrait label reasoning on the user multi-source training data set based on the data set level mapping reasoning component, and obtaining the data set reasoning result of the user multi-source training data set, includes:

[0041] Determining a data subset influence coefficient of each training data subset according to data subset annotation information of each training data subset in the user multi-source training data set, and determining an adjusted training data subset description vector of each training data subset according to the training data subset description vector and the data subset influence coefficient of each training data subset;

[0042] According to the adjusted training data subset description vectors of multiple training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set.

[0043] According to an example of an embodiment of the present disclosure, the step of determining the data subset influence coefficient of each training data subset according to the data subset annotation information of each training data subset in the user multi-source training data set includes:

[0044] For each of the training data subsets in the user multi-source training data set, the following operations are performed:

[0045] Obtaining the number of label types inferred by the first portrait label, and determining an equiprobability distribution of the number of label types;

[0046] determining an error between the data subset annotation information of the training data subset and the equal probability distribution;

[0047] The error is determined as a data subset influence coefficient of the training data subset.

[0048] According to an example of an embodiment of the present disclosure, before performing the following operations on each of the user multi-source training data sets, the method further includes:

[0049] Determine, according to the data point inference results of each of the training data subsets in the user multi-source training data set combination, a representative data subset of each label type obtained by the first portrait label inference in the multiple training data subsets in the user multi-source training data set combination;

[0050] For each of the label types, obtaining an initial description vector of the label type and corresponding initial annotation information, and inertially adjusting the initial description vector according to a representative data subset of the label type to obtain an updated initial description vector of the label type;

[0051] For each of the training data subsets, determining a commonality coefficient between a subset description vector of the training data subset and an updated initial description vector of each of the label types, and determining initial annotation information corresponding to a target updated initial description vector having the highest commonality coefficient as data subset initial annotation information of the training data subset;

[0052] For each of the training data subsets, based on the data subset initial annotation information of the training data subset, inertially adjusting the data subset annotation information of the training data subset to obtain updated data subset annotation information of the training data subset;

[0053] The combining the data subset annotation information of each of the training data subsets in the user multi-source training data set, performing second portrait label reasoning on the user multi-source training data set based on the data set level mapping reasoning component, and obtaining the data set reasoning result of the user multi-source training data set, includes:

[0054] In combination with the updated data subset annotation information of each of the training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set;

[0055] Determining the network loss of the data set portrait label network based on the data point inference results of each of the training data subsets and the errors of the data set annotation information, and the data set inference results of the user multi-source training data set and the errors of the data set annotation information, comprises:

[0056] The network loss of the data set portrait label network is determined based on the errors between the data point inference results of each of the training data subsets and the annotation information of the updated data subset, and the errors between the data set inference results of the user multi-source training data set and the data set annotation information.

[0057] According to an example of an embodiment of the present disclosure, the data point inference result includes the data point inference confidence of the training data subset corresponding to each of the label types; the representative data subsets of each label type obtained by the first portrait label inference in multiple training data subsets of the user multi-source training data set combination are determined based on the data point inference results of each of the training data subsets in the user multi-source training data set combination, including:

[0058] For each of the various tag types, the following operations are performed:

[0059] Determine a set number of target training data subsets corresponding to the label type and having the largest confidence in data point reasoning from the plurality of training data subsets of the user multi-source training data set combination;

[0060] The set number of target training data subsets are used as the set number of representative data subsets of the label type.

[0061] According to an example of an embodiment of the present disclosure, wherein the number of the representative data subsets is x, x>1, the representative data subset based on the tag type, inertially adjusting the initial description vector to obtain an updated initial description vector of the tag type includes:

[0062] Based on a first representative data subset of the label type, inertially adjust the initial description vector to obtain a first hidden initial description vector of the label type;

[0063] Based on the u-th representative data subset of the label type, inertial adjustment of the u-1th hidden initial description vector of the label type is performed to obtain the u-th hidden initial description vector of the label type, 0<u≤x;

[0064] Traversing the representative data subset, obtaining an x-th hidden initial description vector of the label type, and using the x-th hidden initial description vector of the label type as an updated initial description vector of the label type;

[0065] Alternatively, the number of the representative data subsets is multiple, and the inertial adjustment of the initial description vector based on the representative data subset of the tag type to obtain the updated initial description vector of the tag type includes:

[0066] Obtaining a representative subset description vector of each representative data subset of the label type, and determining a mean subset description vector of a plurality of the representative subset description vectors;

[0067] Obtaining a first description vector influence coefficient of the mean subset description vector and a second description vector influence coefficient of the initial description vector;

[0068] According to the first description vector influence coefficient and the second description vector influence coefficient, the mean subset description vector and the initial description vector are fused to obtain an updated initial description vector of the label type;

[0069] The step of inertially adjusting the data subset annotation information of the training data subset based on the data subset initial annotation information of the training data subset to obtain updated data subset annotation information of the training data subset includes:

[0070] Acquire a first annotation influence coefficient of the initial annotation information of the data subset and a second annotation influence coefficient of the annotation information of the data subset;

[0071] According to the first annotation influence coefficient and the second annotation influence coefficient, the data subset initial annotation information and the data subset annotation information are fused to obtain updated data subset annotation information of the training data subset.

[0072] According to another aspect of an embodiment of the present disclosure, there is provided a health care service intelligent matching system, including: a processor; and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the method described in the first aspect above.

[0073] The beneficial effects of this application include at least:

[0074] The embodiment of the present application provides an intelligent matching method and system for health care services based on multi-source big data. When performing portrait label inference on a multi-source data set of a target user, the multi-source data set of the target user is first decomposed to obtain a plurality of data subsets included in the multi-source data set of the target user, and then a description vector is extracted for each data subset to obtain a subset description vector of each data subset. Based on the subset description vector of each data subset, a first portrait label inference is performed on each data subset to obtain a data subset portrait label of each data subset. Thus, based on the data subset portrait label of each data subset, an influence coefficient of each data subset is determined, and each subset description vector is adjusted based on the subset description vector and the influence coefficient of each data subset to obtain an adjusted subset description vector of each data subset. Finally, based on the adjusted subset description vector of each data subset, a second portrait label inference is performed on the multi-source data set of the target user to obtain a user portrait inference result of the multi-source data set of the target user, so as to match health care services based on the user portrait inference result. Because the influence coefficient of adjusting each subset description vector represents the importance of the data subset to the user portrait inference result of the target user multi-source data set, then each adjusted subset description vector obtained by adjustment can more accurately characterize the target user multi-source data set, making the user portrait inference result obtained by inferring the portrait labels of each adjusted subset description vector more accurate, thereby improving the reasonable matching of health care services.

[0075] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The above and other purposes, features and advantages of the embodiments of the present disclosure will become more apparent by describing the embodiments of the present disclosure in more detail in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0077] Figure 1 A schematic diagram of an application scenario provided for an embodiment of the present application.

[0078] Figure 2 A schematic diagram of the implementation process of a method for intelligent matching of health care services based on multi-source big data provided in an embodiment of the present application.

[0079] Figure 3 A schematic diagram of the composition structure of a health care service intelligent matching device provided in an embodiment of the present application.

[0080] Figure 4 A hardware entity diagram of a health care service intelligent matching system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0082] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0083] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict. The terms "first / second / third" involved are merely used to distinguish similar objects and do not represent a specific order for the objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing this application and are not intended to limit this application.

[0085] The intelligent matching method for health care services based on multi-source big data provided in the embodiments of the present application can be applied to Figure 1 In the application environment shown. Among them, each terminal 102 communicates with the health care service intelligent matching system 104 through the network. The data storage system can store the data that the health care service intelligent matching system 104 needs to process. The data storage system can be integrated on the health care service intelligent matching system 104, or it can be placed on the cloud or other network servers. Among them, the health care data can be stored in the local storage of the terminal 102, or it can be stored in the data storage system or the cloud storage associated with the health care service intelligent matching system 104. When service matching is required, the health care service intelligent matching system 104 can obtain multi-source health care data from the terminal 102 local storage, or the data storage system, or the cloud storage. Among them, the terminal 102 can be, but is not limited to, various data source providers, such as health record terminals, medical record terminals, social terminals, Internet of Things devices and portable wearable devices, etc. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The health care service intelligent matching system 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0086] The health care service intelligent matching method based on multi-source big data provided in the embodiment of the present application is applied to the health care service intelligent matching system 104, and specifically includes the following steps:

[0087] Step S110 , decomposing the target user multi-source data set to obtain a plurality of data subsets included in the target user multi-source data set.

[0088] In the embodiment of the present application, the target user multi-source data set is data related to health care services collected from multiple sources, such as user health records, medical records, sensor data, social media data, etc. It can be understood that these data can be updated regularly, and the acquisition of data needs to be carried out within the scope of laws and regulations and with the permission of the user. It can be understood that the data format and content of these multi-source data are not consistent. After acquisition, they need to be normalized, integrated and cleaned to ensure the quality and consistency of the data. In one example, different types of data can be standardized first so that they have the same data format and structure. For example, for text data, natural language processing technology can be used to convert it into a structured form, such as word vectors or text classification, and different types of data can be converted so that they can be compared on the same scale or range. For example, for sensor data, it can be converted into a standardized value or ratio; then, using data integration technology, different types of data are integrated into a unified data warehouse or data lake, which can be achieved through ETL (Extract, Transform, Load) tools or data integration platforms.

[0089] In step S110, after obtaining the target user multi-source data set to be processed, the target user multi-source data set is decomposed to obtain multiple data subsets included in the target user multi-source data set. In the embodiment of the present application, multi-instance learning (MIL) technology is used, wherein the target user multi-source data set is an instance set (also called an instance package or instance bag) in multi-instance learning, and the data subset is the instance (also called a data point) in the instance set. In the embodiment of the present application, when the target user multi-source data set is decomposed, the target user multi-source data set can be decomposed based on the following method to obtain multiple data subsets included in the target user multi-source data set: determine the data type marker contained in the target user multi-source data set, the data type marker includes at least one, and based on the at least one data type marker, the target user multi-source data set is decomposed to obtain multiple data subsets included in the target user multi-source data set. For example, the target user multi-source data set includes 3 type markers, namely, a health record marker, a medical record marker, a sensor marker, a social behavior marker, etc., and each type marker corresponds to a group of data marked under the type. Based on this, it can be understood that the data of each "source" is divided into a corresponding "data cluster", and each data cluster is marked by a type marker. Each type of data can be decomposed directly based on the type marker to obtain data subsets, which can also be called data clusters.

[0090] Step S120, extracting description vectors for each data subset to obtain subset description vectors for each data subset.

[0091] In step S120, for the multiple data subsets included in the target user multi-source data set, description vectors are extracted for each data subset to obtain a subset description vector of each data subset. The description vector described in the embodiment of the present application is a vector representation result of describing feature information obtained by performing feature mining on the target object (such as the target user multi-source data set or data subset). Optionally, description vectors may be extracted for each data subset based on the following method to obtain a subset description vector for each data subset: first-level description vector extraction is performed on each data subset to obtain a first-level hidden subset description vector for each data subset (i.e., an intermediate subset description vector); the first-level hidden subset description vectors of each of the multiple data subsets are merged to obtain a hidden subset description vector array of the target user's multi-source data set (it can be understood that this array is a two-dimensional array, i.e., a matrix); second-level description vector extraction is performed on the hidden subset description vector array to obtain a subset description vector array of the target user's multi-source data set, the subset description vector array including second-level subset description vectors of each data subset, the second-level dimension being smaller than the first-level dimension.

[0092] In an embodiment of the present application, the first-level description vector is first extracted for each data subset to obtain the first-level hidden subset description vector of each data subset. The first level and the second level can be understood as feature representations of different dimensions for the target object. Different levels correspond to different dimensions of the feature space. In one example, the first-level description vector extraction can be implemented based on a pre-trained convolutional neural network. Next, the first-level hidden subset description vectors of each of the multiple data subsets are merged (that is, vector splicing is performed) to obtain an array of hidden subset description vectors of the target user's multi-source data set. Then, the second-level description vector is extracted from the hidden subset description vector array to obtain an array of subset description vectors of the target user's multi-source data set, which includes the second-level subset description vectors of each data subset, and the dimension of the second level is less than the dimension of the first level.

[0093] Specifically, the second-level description vector extraction can be completed based on the pre-debugged backbone network. The backbone network can include a transformer network (Transformer) with a skip connection layer (Skip Connection), which can complete the dimensionality reduction of the description vector and complete the fusion of description vectors between various data subsets in the target user's multi-source data set to increase the accuracy of description vector extraction, so that the obtained subset description vector can more accurately represent the data subset. Based on this, the extracted hidden subset description vectors are merged to obtain the hidden subset description vector array of the target user's multi-source data set, and then the description vector is extracted from the hidden subset description vector array, which can complete the description vector integration between various data subsets in the target user's multi-source data set to increase the accuracy of description vector extraction, reduce the description vector dimension, increase efficiency, and save computing overhead.

[0094] Step S130, performing first portrait label inference on each data subset according to the subset description vector of each data subset, to obtain a data subset portrait label of each data subset.

[0095] In step S130, after obtaining the subset description vector of each data subset, the first portrait label inference can be performed on each data subset according to the subset description vector of each data subset, that is, the corresponding portrait label can be inferred to obtain the data subset portrait label of each data subset. The data subset portrait label indicates the label type corresponding to the corresponding data subset, and the data subset portrait label can also include the data point inference confidence of the data subset corresponding to each label type, where the data point can be understood as a data subset (i.e., an instance). The first portrait label inference can be used for portrait label inference of q label types. In a specific example, the data subset portrait label here can be a probability vector tending to the confidence of each label type, and the number and specific form of the label type are not limited. For example, for a data subset (such as a medication data subset), the label type includes four categories, specifically long-term drug dependence, non-long-term drug dependence, self-medication, and inability to self-medication. The obtained data subset portrait label is (0.8, 0.2, 0.7, 0.3). Then, in this data subset, the portrait label corresponding to the high probability is (long-term drug dependence, self-medication). First, portrait label reasoning can be performed based on the data point level mapping reasoning component of the data set portrait label network (that is, the classification structure at the data point level, which can be a classifier, such as Softmax, fully connected layer). The data set portrait label network can be obtained by pre-debugging, and the data set portrait label network can be a convolutional neural network.

[0096] Step S140, for each data subset, determine the influence coefficient of the data subset based on the data subset portrait label of the data subset, and determine the adjusted subset description vector of the data subset based on the subset description vector and the influence coefficient of the data subset.

[0097] In step S140, after obtaining the data subset portrait label of each data subset, the following operations are performed on each data subset: based on the data subset portrait label of the data subset, the influence coefficient of the data subset is determined, and based on the subset description vector and the influence coefficient of the data subset, the adjusted subset description vector of the data subset is determined. In the present application, the influence coefficient indicates the importance of the data subset (or the subset description vector of the data subset) to the user portrait inference result of the target user multi-source data set, and the influence coefficient is a weight. Then, the subset description vector of the data subset can be weighted according to the influence coefficient to increase the accuracy of the portrait label inference of the target user multi-source data set based on the subset description vector of the data subset.

[0098] Optionally, the influence coefficient of the data subset can be determined based on the data subset portrait label of the data subset based on the following method: obtain the number of label types inferred by the first portrait label, and determine the equiprobability distribution of the number of label types; determine the error between the data subset portrait label of the data subset and the equiprobability distribution; and determine the error as the influence coefficient of the data subset.

[0099] For example, first obtain the number of label types for the first portrait label inference, such as 4 categories, and then determine the equiprobability distribution corresponding to the number of label types. For example, when the number of label types is 4, the equiprobability distribution is (0.25, 0.25, 0.25, 0.25). When the number of label types is 2, the equiprobability distribution is (0.5, 0.5). Then, determine the error between the data subset portrait label and the equiprobability distribution of the data subset, and determine the error as the influence coefficient of the data subset. For example, calculate the relative entropy between the data subset portrait label and the equiprobability distribution, and determine the relative entropy as the influence coefficient of the data subset. For example, the data subset portrait label is (0.8, 0.2, 0.7, 0.3), and the equiprobability distribution is (0.25, 0.25, 0.25, 0.25). The influence coefficient W of the two can be calculated using the following process:

[0100] W= =

[0101] = + + + =1.609.

[0102] Then, the influence coefficient of this data subset is 1.609.

[0103] Optionally, the adjusted subset description vector of the data subset may be determined according to the subset description vector and the influence coefficient of the data subset based on the following method: the subset description vector of the data subset is multiplied by the influence coefficient of the data subset to obtain the adjusted subset description vector of the data subset.

[0104] Step S150, performing a second portrait label inference on the target user multi-source data set based on the adjusted subset description vectors of the multiple data subsets, and obtaining a user portrait inference result of the target user multi-source data set.

[0105] In step S150, after obtaining the adjusted subset description vectors of multiple data subsets in the target user multi-source data set, the target user multi-source data set is subjected to second portrait label reasoning according to the respective adjusted subset description vectors of the multiple data subsets in the target user multi-source data set, and the user portrait reasoning result of the target user multi-source data set is obtained. The user portrait reasoning result is used to indicate the label type corresponding to the target user multi-source data set, and the user portrait reasoning result may also include the reasoning confidence of the target user multi-source data set corresponding to each label type at the data set level (i.e., in the dimension of the instance set). Specifically, the number of label types of the second portrait label reasoning may be the total number of label types of the first portrait label reasoning. The second portrait label reasoning may be performed based on the data set level mapping reasoning component of the data set portrait label network, and the data set portrait label network may be a pre-debugged neural network, such as a convolutional neural network. Specifically, step S150 may include: performing a description vector fusion operation (i.e., aggregating together) on the adjusted subset description vectors of the multiple data subsets to obtain a fused subset description vector; performing a second portrait label reasoning on the target user multi-source data set according to the fused subset description vector to obtain the user portrait reasoning result of the target user multi-source data set.

[0106] Step S160, based on the user portrait inference result, determine the matching health care service in the preset health care service list.

[0107] For example, the preset health care service list contains service tags corresponding to each service. A service tag is a confidence vector, and the multiple elements it contains correspond to the confidence of a service item. In the preset health care service list, the service tag with the greatest vector similarity to the user portrait inference result is determined as the target service tag, and the health care service corresponding to the target service tag is the matched health care service.

[0108] The following is an introduction to the debugging process of the dataset portrait label network. Optionally, the debugging process of the dataset portrait label network includes the following steps:

[0109] Step S210, obtaining a user multi-source training data set combination for debugging a data set portrait label network, the user multi-source training data set combination includes multiple user multi-source training data sets, the user multi-source training data sets carry data set annotation information, the user multi-source training data sets include multiple training data subsets, and the training data subsets carry data subset annotation information; Step S220, based on the data point level mapping reasoning component, respectively perform a first portrait label reasoning on each training data subset in each user multi-source training data set to obtain a data point reasoning result for each training data subset; Step S230, for each user multi-source training data set, perform the following operations: Step S23 1. Combine the data subset annotation information of each training data subset in the user multi-source training data set, and perform the second portrait label inference on the user multi-source training data set based on the data set level mapping reasoning component to obtain the data set inference result of the user multi-source training data set; step S232, determine the network loss of the data set portrait label network according to the data point inference result of each training data subset and the error of the data subset annotation information, and the data set inference result of the user multi-source training data set and the error of the data set annotation information; step S233, adjust the network internal configuration variables of the data set portrait label network according to the network loss to debug the data set portrait label network. Among them, the annotation information is the real information determined in advance, and the network internal configuration variables can include parameters or hyperparameters such as weights, biases, and learning rates.

[0110] In step S210, a user multi-source training data set combination for debugging a data set portrait label network is first obtained, the user multi-source training data set combination includes multiple user multi-source training data sets, each user multi-source training data set includes multiple training data subsets, and the training data subsets are obtained based on decomposing the user multi-source training data set. For example, for each user multi-source training data set, the data type marker of the user multi-source training data set is determined, and the user multi-source training data set is decomposed based on the data type marker to obtain multiple training data subsets included in the user multi-source training data set. Among them, each user multi-source training data set carries data set annotation information, and each training data subset carries data subset annotation information. Specifically, during the initialization process, the data subset annotation information of the training data subset can determine the data set annotation information of the user multi-source training data set corresponding to the training data subset as the data subset annotation information of the training data subset. Because the user multi-source training data set may have a training data subset that is unrelated to the data set annotation information, the data subset annotation information of the training data subset can be regarded as a soft label.

[0111] In step S220, based on the data point level mapping reasoning component of the data set portrait label network, first portrait label reasoning is performed on each training data subset in each user's multi-source training data set to obtain data point reasoning results for each training data subset. Optionally, the data point reasoning results for each training data subset can be obtained based on the following method: obtaining the training data subset description vector of each training data subset in each user's multi-source training data set; based on the data point level mapping reasoning component, according to the training data subset description vector of each training data subset, first portrait label reasoning is performed on each training data subset to obtain data point reasoning results for each training data subset.

[0112] Here, the following operations can be performed on each user's multi-source training data set: obtain the training data subset description vector of each training data subset in the user's multi-source training data set, such as extracting the description vector of each training data subset to obtain the training data subset description vector of each training data subset. The description vector extraction method here can be consistent with the description vector extraction method of the above subset description vector. Based on the training data subset description vector of each training data subset in the user's multi-source training data set, perform the first portrait label inference on each training data subset in each user's multi-source training data set to obtain the data point inference result of each training data subset. The data point inference result represents the label type corresponding to the corresponding training data subset, and the data point inference result can also represent the data point inference confidence of the corresponding training data subset belonging to each label type.

[0113] In step S230, for each user's multi-source training data set, the following operations are performed:

[0114] Step S231, combining the data subset annotation information of each training data subset in the user multi-source training data set, based on the data set level mapping reasoning component, performs second portrait label reasoning on the user multi-source training data set, and obtains the data set reasoning result of the user multi-source training data set. Optionally, the data set reasoning result of the user multi-source training data set can be obtained based on the following method: according to the data subset annotation information of each training data subset in the user multi-source training data set, determine the data subset influence coefficient of each training data subset, and according to the training data subset description vector and the data subset influence coefficient of each training data subset, determine the adjusted training data subset description vector of each training data subset; according to the adjusted training data subset description vectors of multiple training data subsets in the user multi-source training data set, perform second portrait label reasoning on the user multi-source training data set based on the data set level mapping reasoning component, and obtain the data set reasoning result of the user multi-source training data set.

[0115] Here, for each training data subset, the data subset influence coefficient of the training data subset can be determined based on the data subset annotation information and the data point reasoning result of the training data subset, and the data subset influence coefficient indicates the importance of the training data subset to the data set reasoning result of the user's multi-source training data set. Optionally, the data subset influence coefficient of each training data subset can be determined based on the data subset annotation information of each training data subset in the user's multi-source training data set based on the following method: for each training data subset in the user's multi-source training data set, the following operations are performed: the number of label types inferred by the first portrait label is obtained, and an equiprobability distribution of the number of label types is determined; the error between the data subset annotation information of the training data subset and the equiprobability distribution is determined; the error is determined as the data subset influence coefficient of the training data subset.

[0116] Here, when calculating the data subset influence coefficient of the training data subset, first obtain the number of label types inferred by the first portrait label, then determine the equal probability distribution corresponding to the number of label types, and then determine the error between the data point inference result of the training data subset and the equal probability distribution, and determine the error as the data subset influence coefficient of the training data subset. For example, determine the relative entropy between the data point inference result and the equal probability distribution as the data subset influence coefficient of the training data subset.

[0117] According to the data subset influence coefficient, the training data subset description vector of the training data subset is weightedly calculated (that is, the data subset influence coefficient and the training data subset description vector are multiplied to obtain an adjusted training data subset description vector, and then according to the adjusted training data subset description vectors of multiple training data subsets in the user multi-source training data set, the second portrait label inference is performed on the user multi-source training data set based on the data set level mapping reasoning component to obtain the data set reasoning result of the user multi-source training data set. The data set reasoning result is used to indicate the label type corresponding to the user multi-source training data set, and the data set reasoning result can also indicate the data set level reasoning confidence of each label type belonging to the user multi-source training data set.

[0118] Step S232, determine the network loss of the data set portrait label network based on the data point reasoning results of each training data subset and the error of the data set annotation information, and the data set reasoning results of the user's multi-source training data set and the error of the data set annotation information. Optionally, step S232 may further include the following steps: step S2321, for each training data subset, determine the hidden data point level loss (i.e., the loss of the intermediate state) of the data set portrait label network based on the data point reasoning results of the training data subset and the error of the data set annotation information; step S2322, sum up the hidden data point level losses of the data set portrait label network to obtain the data point level loss of the data set portrait label network; step S2323, determine the data set level loss of the data set portrait label network based on the data set reasoning results of the user's multi-source training data set and the error of the data set annotation information; step S2324, fuse the data point level loss and the data set level loss based on the first loss influence coefficient of the data point level loss and the second loss influence coefficient of the data set level loss to obtain the network loss of the data set portrait label network. In steps S2321 and S2323, the hidden data point level loss can be obtained using a cross entropy loss function. The specific values ​​of the first loss influence coefficient and the second loss influence coefficient can be set in advance according to actual needs.

[0119] Step S233, adjusting the network internal configuration variables of the data set portrait label network according to the network loss to debug the data set portrait label network.

[0120] Specifically, the loss can be fed back in the data set portrait label network, and the network internal configuration variables of each component in the data set portrait label network can be adjusted based on the gradient optimization algorithm to complete the debugging of the data set portrait label network. As mentioned above, because the data subset annotation information of the training data subset is a soft label, the data subset annotation information may have interference information, which hinders the debugging convergence of the network. In the embodiment of the present application, the data subset annotation information of the training data subset can also be updated during debugging to complete the denoising of the annotation information to enhance the debugging performance of the data set portrait label network. Assuming that the first portrait label inference is used to perform portrait label inference of p label types, the debugging process of the data set portrait label network provided in the embodiment of the present application, before step S230 and after step S220, can also include:

[0121] Step S10, based on the data point inference results of each training data subset in the user multi-source training data set combination, determine the representative data subset of each label type obtained by the first portrait label inference from the multiple training data subsets of the user multi-source training data set combination; Step S20, for each label type, obtain the initial description vector of the label type and the corresponding initial annotation information, and inertially adjust the initial description vector based on the representative data subset of the label type to obtain the updated initial description vector of the label type; Step S30, for each training data subset, determine the commonality coefficient (that is, the degree of similarity, which can be determined by calculating the distance between vectors, such as Euclidean distance, the closer the distance, the greater the similarity coefficient) between the subset description vector of the training data subset and the updated initial description vector of each label type, and determine the initial annotation information corresponding to the target updated initial description vector with the largest commonality coefficient as the data subset initial annotation information of the training data subset; Step S40, for each training data subset, based on the data subset initial annotation information of the training data subset, inertially adjust the data subset annotation information of the training data subset to obtain the updated data subset annotation information of the training data subset.

[0122] After obtaining the data point inference results of each training data subset in the user multi-source training data set combination in step S220, in step S10, based on the data point inference results of each training data subset in the user multi-source training data set combination, the representative data subset of each label type is determined in the multiple training data subsets of the user multi-source training data set combination. Here, the data point inference results include the data point inference confidence of the training data subset corresponding to each label type, so the representative data subset of each label type can be determined from the multiple training data subsets of the user multi-source training data set combination based on the following method: For each label type, the following operations are performed respectively: From the multiple training data subsets of the user multi-source training data set combination, a set number of target training data subsets corresponding to the label type and having the largest data point inference confidence are determined; the set number of target training data subsets are used as the set number of representative data subsets of the label type. The set number here is less than the number of training data subsets.

[0123] In step S20, for each label type, the initial description vector of the label type can be inertially adjusted based on the representative data subset of the label type to obtain an updated initial description vector of the label type. Here, for each label type, the initial description vector and initial annotation information of the label type can be initialized in advance. The initialized initial description vector can be an all-zero vector, and the label type of the initial annotation information can correspond to the classification label type that can be implemented by the data set portrait label network (the classification label type that can be implemented by the second portrait label reasoning).

[0124] Optionally, the initial description vector provided in the embodiment of the present application also includes an inertia adjustment process, which is a momentum update (Momentum Update), which is an optimization algorithm for accelerating the convergence speed and stability of the gradient descent algorithm. In the traditional gradient descent algorithm, each iteration updates the parameters of the model according to the current gradient direction. However, due to the noise and randomness of the gradient, the update direction of each iteration may fluctuate greatly, resulting in slow and unstable convergence. Inertia adjustment introduces an inertia term (Inertia Term), which adjusts the current update direction according to the gradient direction of the previous iteration. By introducing the inertia term, the momentum update can smooth the change of the gradient and reduce the fluctuation of the update direction, thereby accelerating the convergence speed and improving stability. When the gradient direction is the same as the previous iteration direction, the inertia term will increase the amplitude of the update; when the gradient direction is different from the previous iteration direction, the inertia term will reduce the amplitude of the update to avoid excessive oscillation. Then, step S20 may specifically include the following steps: perform the following operations for each label type: step S201a, based on the first representative data subset of the label type, inertially adjust the initial description vector to obtain the first hidden initial description vector of the label type; step S202a, based on the u-th representative data subset of the label type, inertially adjust the u-1th hidden initial description vector of the label type to obtain the u-th hidden initial description vector of the label type, 0<u≤x; step S203a, traverse the representative data subsets to obtain the x-th hidden initial description vector of the label type, and use the x-th hidden initial description vector of the label type as the updated initial description vector of the label type.

[0125] Steps S201a-~S203a are to adjust the initial description vector once based on each representative data subset to obtain the final updated initial description vector. Taking the first representative data subset based on the label type, inertial adjustment of the initial description vector, and obtaining the first hidden initial description vector of the label type as an example, the process of each inertial adjustment is explained: obtain the representative subset description vector of the first representative data subset; obtain the third description vector influence coefficient of the representative subset description vector and the fourth description vector influence coefficient of the initial description vector; according to the third description vector influence coefficient and the fourth description vector influence coefficient, the representative subset description vector and the initial description vector are fused, and the fusion process here is weighted summation, and the weighted weights are the various influence coefficients, so as to obtain the first hidden initial description vector of the label type. Here, the third description vector influence coefficient and the fourth description vector influence coefficient can be determined in advance according to actual conditions.

[0126] Optionally, in the inertia adjustment of the initial description vector provided in the embodiment of the present application, step S20 may also include the following steps: perform the following operations for each label type: step S201b, obtain the representative subset description vector of each representative data subset of the label type, and determine the mean subset description vector of multiple representative subset description vectors; step S202b, obtain the first description vector influence coefficient of the mean subset description vector and the second description vector influence coefficient of the initial description vector; step S203b, based on the first description vector influence coefficient and the second description vector influence coefficient, fuse the mean subset description vector and the initial description vector to obtain an updated initial description vector of the label type. Here, based on the mean subset description vector of the representative subset description vectors of multiple representative data subsets, the initial description vector is only inertia adjusted once to obtain an updated initial description vector. The first description vector influence coefficient and the second description vector influence coefficient can also be determined in advance according to actual conditions.

[0127] In step S30, the following operations can be performed on each training data subset: determine the commonality coefficient between the subset description vector of the training data subset and the updated initial description vector of each label type, and determine the initial annotation information corresponding to the target updated initial description vector with the largest commonality coefficient as the data subset initial annotation information of the training data subset.

[0128] Optionally, in the inertial adjustment of the data subset annotation information provided in the embodiment of the present application, step S40 may include: performing the following operations for each training data subset: step S41, obtaining the first annotation influence coefficient of the initial annotation information of the data subset and the second annotation influence coefficient of the data subset annotation information; step S42, based on the first annotation influence coefficient and the second annotation influence coefficient, fusing the initial annotation information of the data subset and the data subset annotation information (weighted summation) to obtain the updated data subset annotation information of the training data subset. Here, the data subset annotation information can be inertially adjusted based on the initial annotation information of the data subset. The first annotation influence coefficient and the second annotation influence coefficient can be determined in advance according to actual conditions.

[0129] Optionally, after the data subset annotation information is updated based on steps S10 to S40, the data subset annotation information to be used in step S230 is based on the adjusted data subset annotation information. In other words, step S231 combines the updated data subset annotation information of each training data subset in the user multi-source training data set, performs a second portrait label inference on the user multi-source training data set based on the data set level mapping reasoning component, and obtains the data set inference result of the user multi-source training data set; in step S232, the network loss of the data set portrait label network is determined based on the data point inference result of each training data subset and the error of the updated data subset annotation information, as well as the data set inference result of the user multi-source training data set and the error of the data set annotation information.

[0130] For example, based on the dataset portrait label network, the target user multi-source dataset can be classified into required labels to obtain the corresponding label type. Here, the label type can be a confidence vector of (a, b, c, d, e), where each element represents the confidence of a label type, such as long-term drug dependence, history of allergies, exercise habits, frequent social activities, high consumption habits, etc.

[0131] When the target user multi-source data set is subjected to portrait label inference through the embodiment of the present application, the target user multi-source data set is first decomposed to obtain multiple data subsets included in the target user multi-source data set, and then description vectors are extracted for each data subset to obtain a subset description vector of each data subset, and then first portrait label inference is performed on each data subset based on the subset description vector of each data subset to obtain a data subset portrait label of each data subset; thereby, based on the data subset portrait label of each data subset, an influence coefficient of each data subset is determined, and each subset description vector is weightedly adjusted based on the subset description vector and the influence coefficient of each data subset to obtain an adjusted subset description vector of each data subset, and finally, based on the adjusted subset description vector of each data subset, a second portrait label inference is performed on the target user multi-source data set to obtain a user portrait inference result of the target user multi-source data set. Because the weighted influence coefficient of each subset description vector indicates the importance of the data subset to the user portrait inference result of the target user multi-source data set, the adjusted description vectors of each adjusted subset obtained can more accurately characterize the target user multi-source data set, making the user portrait inference result obtained by inferring the portrait labels of each adjusted subset description vector more accurate, thereby improving the reasonable matching of health care services.

[0132] Based on the same inventive concept, the embodiment of the present application also provides a health care service intelligent matching device for implementing the health care service intelligent matching method based on multi-source big data involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more health care service intelligent matching device embodiments provided below can refer to the limitations of the health care service intelligent matching method based on multi-source big data above, and will not be repeated here.

[0133] In one embodiment, Figure 3 As shown, a health care service intelligent matching device is provided, including:

[0134] A data decomposition module 310 is used to decompose the target user multi-source data set to obtain a plurality of data subsets included in the target user multi-source data set;

[0135] A feature extraction module 320 is used to extract description vectors for each of the data subsets to obtain subset description vectors for each of the data subsets;

[0136] A first reasoning module 330 is used to perform first image label reasoning on each of the data subsets according to the subset description vector of each of the data subsets, so as to obtain a data subset image label of each of the data subsets;

[0137] An influence determination module 340 is used to determine, for each of the data subsets, an influence coefficient of the data subset according to the data subset portrait label of the data subset, and determine an adjusted subset description vector of the data subset according to the subset description vector of the data subset and the influence coefficient, wherein the influence coefficient is used to indicate the importance of the data subset to the user portrait inference result of the target user multi-source data set;

[0138] A second reasoning module 350 is used to perform a second portrait label reasoning on the target user multi-source data set according to the adjusted subset description vectors of the multiple data subsets to obtain a user portrait reasoning result of the target user multi-source data set;

[0139] The service matching module 360 ​​determines matching health care services in a preset health care service list based on the user portrait inference result.

[0140] Each module in the above-mentioned intelligent matching device for health care services can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above modules.

[0141] In one embodiment, a health care service intelligent matching system is provided. The health care service intelligent matching system can be a server, and its internal structure diagram can be as follows: Figure 4 As shown. The health care service intelligent matching system includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the health care service intelligent matching system is used to provide computing and control capabilities. The memory of the health care service intelligent matching system includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the health care service intelligent matching system is used to store data including multi-source health care data. The input / output interface of the health care service intelligent matching system is used to exchange information between the processor and an external device. The communication interface of the health care service intelligent matching system is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a health care service intelligent matching method based on multi-source big data is implemented.

[0142] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the intelligent matching system for health care services to which the present application scheme is applied. The specific intelligent matching system for health care services may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0143] In one embodiment, a health care service intelligent matching system is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0144] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0145] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0146] It should be noted that the object information (including but not limited to the object's device information, corresponding personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0147] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0148] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A health care service intelligent matching method based on multi-source big data, characterized in that: The method comprises: Decomposing a target user multi-source data set to obtain a plurality of data subsets included in the target user multi-source data set; Extracting description vectors for each of the data subsets to obtain subset description vectors for each of the data subsets; According to the subset description vector of each of the data subsets, first image label inference is performed on each of the data subsets to obtain a data subset image label of each of the data subsets; For each of the data subsets, determining an influence coefficient of the data subset according to the data subset portrait label of the data subset, and determining an adjusted subset description vector of the data subset according to the subset description vector of the data subset and the influence coefficient, wherein the influence coefficient is used to indicate the importance of the data subset to the user portrait inference result of the target user multi-source data set; Performing a second portrait label inference on the target user multi-source data set according to the adjusted subset description vectors of the multiple data subsets to obtain a user portrait inference result of the target user multi-source data set; Based on the user portrait inference result, determining matching health care services in a preset health care service list; Wherein, determining the influence coefficient of the data subset according to the data subset portrait label of the data subset includes: Obtaining the number of label types inferred by the first portrait label, and determining an equiprobability distribution of the number of label types; Determining an error between a data subset image label of the data subset and the equal probability distribution; determining the error as an influence coefficient of the data subset; The step of determining an adjusted subset description vector of the data subset based on the subset description vector of the data subset and the influence coefficient includes: Multiplying the subset description vector of the data subset by the influence coefficient of the data subset to obtain the adjusted subset description vector of the data subset; The performing second portrait label inference on the target user multi-source data set based on the adjusted subset description vectors of the multiple data subsets to obtain a user portrait inference result of the target user multi-source data set includes: Performing a description vector fusion operation on the adjusted subset description vectors of the multiple data subsets to obtain a fused subset description vector; According to the fused subset description vector, a second portrait label inference is performed on the target user multi-source data set to obtain a user portrait inference result of the target user multi-source data set.

2. The method according to claim 1, characterized in that The extracting description vectors from each of the data subsets to obtain subset description vectors of each of the data subsets includes: Extracting the first-level description vector of each of the data subsets to obtain the first-level hidden subset description vector of each of the data subsets; Merging the first-level hidden subset description vectors of the respective multiple data subsets to obtain an array of hidden subset description vectors of the target user multi-source data set; A second-level description vector extraction is performed on the hidden subset description vector array to obtain a subset description vector array of the target user multi-source data set, wherein the subset description vector array includes the second-level subset description vectors of each of the data subsets, and the dimension of the second level is smaller than the dimension of the first level.

3. The method according to claim 1, characterized in that The first portrait label reasoning is performed based on a data point level mapping reasoning component of a data set portrait label network, and the second portrait label reasoning is performed based on a data set level mapping reasoning component of the data set portrait label network; The method further comprises: Acquire a user multi-source training data set combination for debugging the data set portrait label network, wherein the user multi-source training data set combination includes multiple user multi-source training data sets, the user multi-source training data sets carry data set annotation information, and the user multi-source training data sets include multiple training data subsets, and the training data subsets carry data subset annotation information; Based on the data point level mapping reasoning component, first portrait label reasoning is performed on each of the training data subsets in each of the user multi-source training data sets to obtain data point reasoning results for each of the training data subsets; For each of the user multi-source training data sets, the following operations are performed: In combination with the data subset annotation information of each of the training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set; Determine the network loss of the data set portrait label network according to the errors between the data point inference results of each of the training data subsets and the data set annotation information, and the errors between the data set inference results of the user multi-source training data set and the data set annotation information; The network internal configuration variables of the data set portrait label network are adjusted according to the network loss to debug the data set portrait label network.

4. The method according to claim 3, characterized in that Determining the network loss of the data set portrait label network based on the data point inference results of each of the training data subsets and the errors of the data set annotation information, and the data set inference results of the user multi-source training data set and the errors of the data set annotation information, comprises: For each of the training data subsets, determining the hidden data point level loss of the data set portrait labeling network according to the data point inference result of the training data subset and the error of the data subset annotation information; Sum the hidden data point level losses of the data set portrait label network to obtain the data point level loss of the data set portrait label network; Determine a dataset level loss of the dataset portrait label network according to the dataset inference result of the user multi-source training dataset and the error of the dataset annotation information; According to the first loss influence coefficient of the data point level loss and the second loss influence coefficient of the data set level loss, the data point level loss and the data set level loss are fused to obtain the network loss of the data set portrait label network; The data point level mapping reasoning component performs first portrait label reasoning on each of the training data subsets in each of the user multi-source training data sets to obtain data point reasoning results for each of the training data subsets, including: Obtaining a training data subset description vector of each of the training data subsets in each of the user multi-source training data sets; Based on the data point level mapping reasoning component, according to the training data subset description vector of each training data subset, first image label reasoning is performed on each training data subset to obtain data point reasoning results of each training data subset; The combining the data subset annotation information of each of the training data subsets in the user multi-source training data set, performing second portrait label reasoning on the user multi-source training data set based on the data set level mapping reasoning component, and obtaining the data set reasoning result of the user multi-source training data set, includes: Determining a data subset influence coefficient of each training data subset according to data subset annotation information of each training data subset in the user multi-source training data set, and determining an adjusted training data subset description vector of each training data subset according to the training data subset description vector and the data subset influence coefficient of each training data subset; According to the adjusted training data subset description vectors of multiple training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set.

5. The method according to claim 4, characterized in that The step of determining the data subset influence coefficient of each training data subset according to the data subset annotation information of each training data subset in the user multi-source training data set includes: For each of the training data subsets in the user multi-source training data set, the following operations are performed: Obtaining the number of label types inferred by the first portrait label, and determining an equiprobability distribution of the number of label types; determining an error between the data subset annotation information of the training data subset and the equal probability distribution; The error is determined as a data subset influence coefficient of the training data subset.

6. The method according to claim 3, characterized in that Before performing the following operations on each of the user multi-source training data sets, the method further includes: Determine, according to the data point inference results of each of the training data subsets in the user multi-source training data set combination, a representative data subset of each label type obtained by the first portrait label inference in the multiple training data subsets in the user multi-source training data set combination; For each of the label types, obtaining an initial description vector of the label type and corresponding initial annotation information, and inertially adjusting the initial description vector according to a representative data subset of the label type to obtain an updated initial description vector of the label type; For each of the training data subsets, determining a commonality coefficient between a subset description vector of the training data subset and an updated initial description vector of each of the label types, and determining initial annotation information corresponding to a target updated initial description vector having the highest commonality coefficient as data subset initial annotation information of the training data subset; For each of the training data subsets, based on the data subset initial annotation information of the training data subset, inertially adjusting the data subset annotation information of the training data subset to obtain updated data subset annotation information of the training data subset; The combining the data subset annotation information of each of the training data subsets in the user multi-source training data set, performing second portrait label reasoning on the user multi-source training data set based on the data set level mapping reasoning component, and obtaining the data set reasoning result of the user multi-source training data set, includes: In combination with the updated data subset annotation information of each of the training data subsets in the user multi-source training data set, a second portrait label inference is performed on the user multi-source training data set based on the data set level mapping inference component to obtain a data set inference result of the user multi-source training data set; Determining the network loss of the data set portrait label network based on the data point inference results of each of the training data subsets and the errors of the data set annotation information, and the data set inference results of the user multi-source training data set and the errors of the data set annotation information, comprises: The network loss of the data set portrait label network is determined based on the errors between the data point inference results of each of the training data subsets and the annotation information of the updated data subset, and the errors between the data set inference results of the user multi-source training data set and the data set annotation information.

7. The method according to claim 6, characterized in that The data point inference result includes the data point inference confidence of the training data subset corresponding to each of the label types; the representative data subsets of each label type obtained by the first portrait label inference are determined from multiple training data subsets in the user multi-source training data set combination based on the data point inference results of each of the training data subsets in the user multi-source training data set combination, including: For each of the various tag types, the following operations are performed: Determine a set number of target training data subsets corresponding to the label type and having the largest confidence in data point reasoning from the plurality of training data subsets of the user multi-source training data set combination; The set number of target training data subsets are used as the set number of representative data subsets of the label type.

8. The method according to claim 6, characterized in that The number of the representative data subsets is x, where x>1, and the representative data subsets based on the tag type, inertially adjusting the initial description vector to obtain an updated initial description vector of the tag type, includes: Based on a first representative data subset of the label type, inertially adjust the initial description vector to obtain a first hidden initial description vector of the label type; Based on the u-th representative data subset of the label type, inertial adjustment of the u-1th hidden initial description vector of the label type is performed to obtain the u-th hidden initial description vector of the label type, 0<u≤x; Traversing the representative data subset, obtaining an x-th hidden initial description vector of the label type, and using the x-th hidden initial description vector of the label type as an updated initial description vector of the label type; Alternatively, the number of the representative data subsets is multiple, and the inertial adjustment of the initial description vector based on the representative data subset of the tag type to obtain the updated initial description vector of the tag type includes: Obtaining a representative subset description vector of each representative data subset of the label type, and determining a mean subset description vector of a plurality of the representative subset description vectors; Obtaining a first description vector influence coefficient of the mean subset description vector and a second description vector influence coefficient of the initial description vector; According to the first description vector influence coefficient and the second description vector influence coefficient, the mean subset description vector and the initial description vector are fused to obtain an updated initial description vector of the label type; The step of inertially adjusting the data subset annotation information of the training data subset based on the data subset initial annotation information of the training data subset to obtain updated data subset annotation information of the training data subset includes: Acquire a first annotation influence coefficient of the initial annotation information of the data subset and a second annotation influence coefficient of the annotation information of the data subset; According to the first annotation influence coefficient and the second annotation influence coefficient, the data subset initial annotation information and the data subset annotation information are fused to obtain updated data subset annotation information of the training data subset.

9. A health care service intelligent matching system, characterized in that: include: processor; and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the method as claimed in any one of claims 1 to 8.