Multi-label disease data classification model construction method based on representation learning

Abstract

The invention belongs to the technical field of medical data mining and artificial intelligence, and discloses a multi-label disease data classification model construction method based on representation learning, which comprises the following steps of: perfecting related information on a disease side and a patient side to form a medical knowledge graph, and converting entities and relationships contained in the medical knowledge graph into embedded vectors; designing a hierarchical attention mechanism to extract high-order structures and semantic information of nodes in the knowledge graph, and generating embedded vectors of target nodes in different layers; integrating the embedded vectors of the patient nodes and the disease nodes of different layers, and performing data classification; and constructing a knowledge graph loss function and a disease data classification loss function, and designing dynamic sampling and loss weighting based on node degrees by using a self-adversarial negative sampling strategy to complete model construction. According to the method, various medical data can be efficiently utilized, the complex relation between diseases is fully mined, the problems of data sparsity and class imbalance are solved, and the accuracy and generalization ability of the model on disease data classification are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical data mining and artificial intelligence technology, and specifically relates to a method for constructing a multi-label disease data classification model based on representation learning. Background Art

[0002] Data mining is the process of extracting potential and valuable knowledge from a large amount of data. In the medical field, the use of big data analysis technology for data classification has become a research hotspot. Multi-morbidity is defined as two or more diseases that have a long-term impact on the health of the body. It is an increasingly serious public health challenge facing mankind. Multi-label disease data classification takes into account the correlation between diseases and can classify disease data based on electronic health records (EHR), which is convenient for doctors to conduct further judgment and analysis based on model recognition results, and has gradually become a research hotspot.

[0003] However, the complex association mechanisms and common risk factors between diseases make accurate classification of disease data face many challenges. The quality and sparsity issues of medical data still exist, resulting in insufficient or missing data available for analysis, which increases the difficulty of classification; and traditional disease data classification models often find it difficult to fully capture the complex relationship patterns between medical data nodes, resulting in limited model classification accuracy; in addition, the imbalanced distribution of data categories is a major obstacle in the training of disease data classification models. The data of certain diseases or specific patient groups are relatively scarce, making it difficult for the model to fully learn the characteristics of these minority categories during training, which can easily cause the model to overfit or underfit. Summary of the invention

[0004] To solve the above technical problems, the present invention provides a method for constructing a multi-label disease data classification model based on representation learning, which efficiently utilizes various types of medical data, fully explores the complex relationships between diseases, overcomes the problems of data sparsity and category imbalance, and improves the accuracy and generalization ability of the model in classifying disease data.

[0005] The method for constructing a multi-label disease data classification model based on representation learning described in the present invention comprises the following steps: S1. Using medical data sets, we can perform semantic analysis on medical text knowledge on the disease side, improve the internal connections between various medical concepts mentioned in EHR, and improve personal basic information, examination index information, and current medical history data on the patient side to form a complete medical knowledge graph; S2, mapping the entities and relations contained in the medical knowledge graph into the embedding space, and using the lineaRE model to convert the entities and relations into embedding vectors; S3. Adopt graph convolution operation based on attention mechanism, design hierarchical attention mechanism, extract high-order structure and semantic information of nodes in knowledge graph, and generate embedding vectors of target nodes at different layers; S4, integrating the embedding vectors of patient nodes and disease nodes at different layers, and performing data classification by calculating the score obtained by multiplying the embedding vectors between patient nodes and disease nodes; S5. Construct the knowledge graph loss function and disease data classification loss function, use the self-adversarial negative sampling strategy, design dynamic sampling and loss weighting based on node degree, and complete the construction of the disease data classification model.

[0006] Furthermore, S2 is specifically as follows: embedding is regarded as a linear regression task, and the head entity h and the tail entity t are mapped to low-dimensional vector representations , the embedding vector of relation r is given by Indicates that each dimension of the relationship is represented as a straight line in a rectangular coordinate system; For any given triple , the relationship between the head entity h, relation r and tail entity t is shown as follows: , in, are two weight vectors, are bias vectors that satisfy , represents a set of k-dimensional real numbers; the symbol stands for element-wise product; For each triple , construct the scoring function as shown below: , in, represents the 1-norm, that is, the sum of the absolute values ​​of the elements of the vector; if If the score of is low, it means that the possibility of the relationship r between the head entity h and the tail entity t is high, otherwise it is low.

[0007] Furthermore, in the embedding space, the relation r has different connection modes to meet the corresponding requirements, as follows: When the relationship r is a symmetric mode, that is, the relationship between the two entities is mutual, if A has this relationship with B, then B should also have this relationship with A. In this case, ; When the relationship r is an antisymmetric mode, that is, if A and B have this relationship, then B and A do not have this relationship. At this time, it is required that in any dimension i, ; When the relation r is in reverse mode, that is, if there is , then there exists , at this time it is required ; When the relation r is in composite mode, that is, if there is a relation and ,but , at this time it is required .

[0008] Furthermore, S3 is specifically: S31. Considering the differences in the impact of different entities on the target entity, the attention mechanism is used to calculate the weight scores of different entities, as shown in the following formula: , in, is an exponential function ; is the temperature parameter, which is used to control the distribution range of attention scores; S32, realize the aggregation of neighbor information centered on the head entity h; S321. For each order of neighbor nodes, a hierarchical attention mechanism is designed to dynamically adjust the weights of neighbor nodes of different orders; , , in, is the hierarchical attention score; is the hierarchical weight coefficient; is a learnable parameter used to indicate the importance of the t-th order neighbor node; Indicates when The learnable parameters when ; S322, triples consisting of the head entity h and its first-order neighbors Traverse and fully integrate the embedding information of the head entity h itself and the embedding information of its first-order neighbor nodes to obtain the first-order aggregation vector : , , in, is a trainable weight matrix, is the dimension of the input entity embedding vector h, t, is the dimension of the output vector, and Represents the initial embedding vectors of the head entity h and the tail entity t; A set of first-order triples representing the head entity h, represents the first-order neighbor node; The function is used to perform nonlinear transformation on the aggregated vector to avoid gradient disappearance, where x represents the input value of the function; is the cumulative distribution function of the standard normal distribution; S323, adopting a step-by-step iterative method to gradually incorporate higher-order neighbor information, and realize aggregation of high-order neighbor information centered on the head entity h; ; in, is the central node h The order embedding expression, is the central node h The order embedding expression, The central node h is The first-order neighbor node embedding on the .

[0009] Further, S4 is specifically: After the graph convolution network of the 0th layer is propagated, the disease nodes and patient nodes in the medical knowledge graph are all obtained from the 0th layer to the The embedding vector of the layer is integrated to obtain the complete embedding vector of the patient node p and the disease node d and : , in, , Represents the patient node p and disease node d in the The embedding vector of the layer; the prediction score between the two is calculated by the dot product operation of the vectors: , in, is transposed.

[0010] Furthermore, S5 is specifically: S51. Combined with the self-adversarial negative sampling strategy, dynamic sampling and loss weighting based on node degree are designed to construct the knowledge graph loss function; , , in, yes Negative sampling samples, i represents the index of all negative samples, , is the scoring function for negative triples, j represents the index of the current negative sample, is the current negative sample sampling probability, is the temperature parameter; Indicates the degree of the sample node, that is, the number of relationships that connect the node to other nodes; Represents the maximum degree in the knowledge graph; is a hyperparameter used to adjust the influence strength of degree; is a hyperparameter, which is expressed by Dynamically adjust the sampling distribution according to the change of Constructing knowledge graph loss function , to optimize the representation of entities and relations in vector space: , in, is a fixed interval; is a smoothing factor to prevent the denominator from being 0; n represents the number of negative samples corresponding to each positive sample; is the activation function; S52. Constructing disease data classification loss function : , in, is a positive sample, is a negative sample, is the sigmoid function; S53. Construct a comprehensive loss function for multi-label disease data classification; as shown in the following formula: , in, Represents a node collection size; is the regularization coefficient, which is used to control the strength of the regularization term; Represents the L2 norm square of all node embedding vectors; by minimizing the comprehensive loss function The model is trained and optimized to achieve the best accuracy in data recognition and classification.

[0011] The beneficial effects described in the present invention are as follows: the method described in the present invention comprehensively considers four relationship connection modes: symmetry, antisymmetry, inversion, and combination, so that the model can more accurately understand and model the intrinsic connection between diseases and diagnostic basis, patients and examination indicators, diseases and their comorbidities, etc., enhances the accuracy of embedding vector expression of entities and relationships, and improves the ability to capture semantic structure and intrinsic logic in medical knowledge graphs; proposes a new node aggregation method based on graph convolutional networks and hierarchical attention mechanisms to dynamically capture useful information between entities at different layers, integrate four relationship connection modes, and improve the quality of node embedding vector representation; in addition, a self-adversarial negative sampling strategy is adopted, and dynamic sampling and loss weighting based on node degree are designed to optimize the problems of overfitting of multi-sample data classification and under-learning of few-sample data classification under medical knowledge graphs, solve the imbalance of data, and improve the overall performance of disease data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the process of the method of the present invention; Figure 2 It is a flowchart for building a medical knowledge graph; Figure 3 It is an example diagram of a medical knowledge graph containing different relational connection patterns; Figure 4 It is a high-order neighbor information aggregation graph based on the medical knowledge graph; Figure 5 This is a schematic diagram of the MLDP-KGRL multi-label disease data classification model structure; Figure 6 It is a graph showing the experimental results of whether the polymerization mechanism of the present invention is adopted; Figure 7 To draw the experimental result diagram for whether to adopt the four relationship connection modes; Figure 8 This is a graph showing the experimental results of whether the negative sampling strategy of the present invention is adopted. DETAILED DESCRIPTION

[0013] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0014] like Figure 1 As shown, the method for constructing a multi-label disease data classification model based on representation learning described in the present invention includes the following steps: S1. Using medical data sets, we can perform semantic analysis on medical text knowledge on the disease side, improve the internal connections between various medical concepts mentioned in EHR, and improve personal basic information, examination index information, and current medical history data on the patient side to form a complete medical knowledge graph; S2. Map the entities and relations contained in the medical knowledge graph into the embedding space, and use the lineaRE model (paper: LineaRE: Simple but Powerful Knowledge Graph Embedding for LinkPrediction) to convert the entities and relations into embedding vectors; S3. Adopt graph convolution operation based on attention mechanism, design hierarchical attention mechanism, extract high-order structure and semantic information of nodes in knowledge graph, and generate embedding vectors of target nodes at different layers; S4, integrating the embedding vectors of patient nodes and disease nodes at different layers, and performing data classification by calculating the score obtained by multiplying the embedding vectors between patient nodes and disease nodes; S5. Construct the knowledge graph loss function and disease data classification loss function, use the self-adversarial negative sampling strategy, design dynamic sampling and loss weighting based on node degree, and build a disease data classification model.

[0015] This embodiment constructs a medical knowledge graph based on the public dataset MIMIC-IV of real hospitals provided by the National Institutes of Health of the United States.

[0016] like Figure 2 As shown in Figure 1, S1 constitutes a complete medical knowledge graph, as follows: S11. Extract the last medical record of the patient with diabetes, which contains the latest diagnosis and treatment information. Integrate according to key fields such as patient identifier (patient ID), consultation time, etc. to obtain the patient's basic personal information, living habits, current medical history, patient symptoms, examination indicators, etc.; S12. For each unique entity (including five types of entities: current medical history, patient symptoms, medical examination indicators, lifestyle information, and personal basic information), add medical information related to the entity (such as comorbidity information, high-risk populations, common causes, and diagnostic basis) to inject medical knowledge into the knowledge graph; S13. Numerical entities such as age, examination indicators, and body mass index (BMI) are classified according to medical standards. For example, BMI is divided into four levels: underweight, normal, overweight, and obese. The medical significance of each level is clearly defined. For example, elevated alanine aminotransferase may lead to chronic hepatitis. S14, convert the identified entities and relations into triples, such as (patient ID, suffering from,type 2 diabetes), (diabetes, may lead to,cardiovascular and cerebrovascular diseases), (patient ID, BMI value, obesity); S15. Disambiguate and deduplicate various entities to ensure that each entity with the same meaning has a unique representation; S16. Obtaining medical knowledge graph , where patient p, disease d, and nodes e in EHR and medical text knowledge all belong to the entity set ; Represents a set of relations in a medical knowledge graph; represents the set of triples in the medical knowledge graph; the specific amount of data contained in the medical knowledge graph is shown in Table 1, and the example of the medical knowledge graph is shown in Figure 3 shown.

[0017] Table 1 The amount of data contained in the medical knowledge graph

[0018] In S2, for each triple , construct the scoring function, as shown in formula (1): (1), in, are two weight vectors, are bias vectors that satisfy , represents a set of k-dimensional real numbers; represents the 1-norm, that is, the sum of the absolute values ​​of the elements of the vector; if If the score of is low, it means that the possibility of relationship r between the head entity h and the tail entity t is high, otherwise it is low.

[0019] In the embedding space, the relation r has different connection patterns; and in the medical knowledge graph, such as Figure 3 As shown, the relationships between different entities also belong to different relationship connection modes.

[0020] When the relationship r belongs to different relationship connection modes, the requirements to be met are also different. For example, when the relationship r is a symmetric mode, that is, the relationship between the two entities is mutual, if A has this relationship with B, then B should also have this relationship with A. In this case, the requirement is ; When the relationship r is an antisymmetric mode, that is, if A and B have this relationship, then B and A do not have this relationship. At this time, it is required that in any dimension i, ; When the relation r is in reverse mode, that is, if there is , then there exists , at this time it is required ; When the relationship is in composite mode, that is, if there is a relationship and , then it can be inferred that , at this time it is required Through the above method, four different relationship connection modes are efficiently modeled, which helps the model obtain more accurate semantic and structural information in the knowledge graph.

[0021] S3 is specifically:

[0022] S31. In order to effectively obtain the information stored in different triples, considering the different effects of different entities on the target entity, the attention mechanism is used to calculate the weight scores of different entities, as shown in the following formula: (2), in, is an exponential function ; is a temperature parameter used to control the distribution range of attention scores; a larger The value will make the attention scores more evenly distributed, while smaller The value will make the attention score more focused; in this specific embodiment, The weight score is calculated by calculating the distance between the head and tail entities in the complex space and using the principle that the lower the distance score, the higher the weight , thereby indicating which neighbor nodes should be given more attention.

[0023] In the calculation of attention scores, in order to ensure the comparability of weight scores, especially when dealing with multiple neighbors of the same head entity, normalization is used, as shown in formula (3): (3), Here, Represents all first-order neighbor nodes with the head entity h as the central node.

[0024] S32, realize the aggregation of neighbor information centered on the head entity h; S321. In order to more effectively integrate multi-order neighbor information, a hierarchical attention mechanism is designed: Specifically, for each order of neighbor nodes, a hierarchical attention mechanism is designed to dynamically adjust the weights of neighbor nodes of different orders; (4), (5), in, is the hierarchical attention score; is the hierarchical weight coefficient; is a learnable parameter used to indicate the importance of the t-th order neighbor node; Indicates when The learnable parameters when ; S322, triples consisting of the head entity h and its first-order neighbors Traverse and fully integrate the embedding information of the head entity h itself and the embedding information of its first-order neighbor nodes to obtain the first-order aggregation vector : (6), (7), in, is a trainable weight matrix, is the dimension of the input entity embedding vector h, t, is the dimension of the output vector, and Represents the initial embedding vectors of the head entity h and the tail entity t; A set of first-order triples representing the head entity h, represents the first-order neighbor node; The function is used to perform nonlinear transformation on the aggregated vector to avoid gradient disappearance, where x represents the input value of the function, i.e., the embedded vector in formula (6); is the cumulative distribution function of the standard normal distribution; this aggregation method can more effectively integrate the information of the head entity and its neighboring nodes, improving the performance of the model in prediction tasks; S323, realize the aggregation of high-order neighbor information centered on the head entity h, such as Figure 4 As shown in the figure, a step-by-step iterative method is adopted to gradually incorporate higher-order neighbor information as the aggregation process progresses. As shown in formula (8), the central node h Order embedding expression is based on its Embedded expression of order And the central node h is First-order neighbor node embedding on To calculate: (8).

[0025] S4, integrate the embedding vectors of patients and diseases at different layers, and calculate the scores between patient and disease nodes for data classification, such as Figure 5 Specifically, after After the graph convolution network of layers is propagated, each node will obtain The embedding vectors of the layers capture the characteristics of nodes at different network depths: (9), in, , Represents the patient node p and disease node d in the The embedding vector of the layer, in this specific embodiment, The embedding vectors of each layer are combined into a complete embedding vector using formula (9). , we can get the complete embedding vector of patient p and disease d and As shown in formula (10), the prediction score between the two is calculated by the dot product operation of the vectors: (10).

[0026] S5 is as follows:

[0027] S51. Combined with the self-adversarial negative sampling strategy, dynamic sampling and loss weighting based on node degree are designed to construct the knowledge graph loss function.

[0028] S511. In the knowledge graph, a triple consisting of nodes connected by relationships is regarded as a true triple, in which the head and tail entities constitute positive samples. In order to train the model, the head and tail entities in the true triple are usually randomly replaced to generate negative samples. However, the randomly generated negative samples may contain potential positive samples. The present invention introduces a self-adversarial negative sampling method to dynamically select negative samples that are more likely to mislead the model according to the current performance of the model. Specifically, it can be seen from formula (1) that negative samples with higher scores mean that the model believes that they are less likely to be potential true samples. In order to avoid overfitting and negative sampling of the model under long-tail data, dynamic sampling and loss weighting based on node degree are designed to alleviate the impact of data imbalance to a certain extent and improve the accuracy and generalization ability of the model; (11), (12), in, yes Negative sampling samples, i represents the index of all negative samples, , is the scoring function for negative triples, j represents the index of the current negative sample, is the current negative sample sampling probability, is the temperature parameter. ; Indicates the degree of the sample node, that is, the number of relationships that connect the node to other nodes; Represents the maximum degree in the knowledge graph; is a hyperparameter used to adjust the influence strength of the degree, in this embodiment; is a hyperparameter, which is expressed by The sampling distribution is adjusted dynamically according to the changes in . When the node degree is high, the negative sampling distribution is more concentrated, reducing the overfitting of the model; when the node degree is low, the negative sampling distribution is more uniform, thus avoiding underfitting of the model.

[0029] S512. Construct knowledge graph loss function , to optimize the representation of entities and relations in the vector space. The specific form of the knowledge graph loss function is: (13), in, is a fixed interval (margin). In this embodiment, ; is a smoothing factor to prevent the denominator from being 0. In this embodiment, , n represents the number of negative samples corresponding to each positive sample. In this embodiment, , is the activation function.

[0030] S52. Construct a disease data classification loss function. Specifically, as shown in formula (14), when the probability that patient p suffers from disease d is higher, the disease data classification loss function The smaller: (14), in, is a positive sample, is a negative sample, is the sigmoid function.

[0031] S53, construct a comprehensive loss function for multi-label disease data classification. Specifically, as shown in formula (15), the present invention defines a comprehensive loss function , which combines the knowledge graph loss function , disease data classification loss function And the regularization term to prevent overfitting: (15), in, Represents a node collection size; is the regularization coefficient, which is used to control the strength of the regularization term; represents the L2 norm square of all node embedding vectors. By minimizing this comprehensive loss function The model is trained and optimized to achieve the best accuracy in multi-label disease data classification under the dataset of medical knowledge graph.

[0032] A simulation test was conducted on the multi-label disease data classification model based on representation learning. The test was based on the knowledge graph dataset generated in step S1. The results are shown in Table 2, where the best experimental results are indicated by bold, the suboptimal experimental results are indicated by underscores, and the gap between the best and the suboptimal is indicated by the percentage in brackets.

[0033] The experimental evaluation indicators are as follows: 1) Precision@n (p@n): For each patient, calculate the ratio of the number of actual positive samples in the first n prediction items to the total number of predictions; 2) Recall@n (R@n): For each patient, calculate the ratio of the number of positive samples in the first n prediction items to the number of patients who actually have the disease; 3) F1 score@n (F1@n): The F1 score is the harmonic mean of precision and recall, and is used to comprehensively measure the performance of the two; 4) Normalized discounted cumulative gain@n (NDCG@n): An indicator that considers the importance of the predicted item position is used in the present invention to measure the ability of the model to place the disease that the patient actually suffers from at a higher position in the recommendation list.

[0034] The experimental results in Table 2 show that the multi-label disease data classification model MLDP-KGRL in the present invention has strong robustness. Under the same imbalanced data, the various performance indicators of the present invention have the best results compared with other comparative models. And four different relationship connection modes are fully considered to optimize the accuracy of entity embedding vectors and relationship embedding vectors. At the same time, the advantages of the knowledge graph are fully utilized to incorporate the comorbidity information, high-risk populations, common causes, diagnostic basis, etc. of the disease into the knowledge graph, reducing the sparsity of the data. Compared with the method of only constructing a disease-patient bipartite graph or simply constructing a knowledge graph through the patient's EHR, the model of the present invention can capture richer information.

[0035] Table 2 Comparison of model performance

[0036] Figure 6 The experimental comparison results of whether the aggregation method described in the present invention is adopted in a specific embodiment are shown. The experimental results show that the adoption of the aggregation method of the present invention can significantly improve the accuracy of the model of the present invention. Specifically, the present invention designs an aggregation mechanism for embedding vectors of its own nodes and high-order neighbor nodes, which can self-learn the weight parameters in the aggregation mechanism during the model training process, thereby achieving more accurate vector aggregation.

[0037] Figure 7The experimental comparison results of whether or not to use four relational connection modes for embedding vector generation in a specific embodiment are shown. The experimental results show that the method of using four relational connection modes can significantly improve the accuracy of the model of the present invention compared to using only two relational connection modes, symmetric and asymmetric. Specifically, the present invention considers four different relational connection modes in the medical knowledge graph, optimizes the modeling capability, improves the accuracy of the embedding vector expression of entities and relationships, and thus improves the accuracy of disease data classification.

[0038] Figure 8 The experimental comparison results of whether to use the dynamic sampling based on node degree and the loss weighting method in a specific embodiment are shown. The experimental results show that the method designed by the present invention can significantly improve the classification accuracy of the model of the present invention compared with random negative sampling.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A method for constructing a multi-label disease data classification model based on representation learning, characterized in that: The following steps are involved: S1. Using medical data sets, we can perform semantic analysis on medical text knowledge on the disease side, improve the internal connections between various medical concepts mentioned in EHR, and improve personal basic information, examination index information, and current medical history data on the patient side to form a complete medical knowledge graph; S2, mapping the entities and relations contained in the medical knowledge graph into the embedding space, and using the lineaRE model to convert the entities and relations into embedding vectors; S3. Adopt graph convolution operation based on attention mechanism, design hierarchical attention mechanism, extract high-order structure and semantic information of nodes in knowledge graph, and generate embedding vectors of target nodes at different layers; S4, integrating the embedding vectors of patient nodes and disease nodes at different layers, and performing data classification by calculating the score obtained by multiplying the embedding vectors between patient nodes and disease nodes; S5. Construct the knowledge graph loss function and disease data classification loss function, use the self-adversarial negative sampling strategy, design dynamic sampling and loss weighting based on node degree, and complete the construction of the disease data classification model.

2. According to the method for constructing a multi-label disease data classification model based on representation learning in claim 1, it is characterized in that: S2 is specifically: Embedding is regarded as a linear regression task, and the head entity h and the tail entity t are mapped to low-dimensional vector representations. , the embedding vector of relation r is given by Indicates that each dimension of the relationship is represented as a straight line in a rectangular coordinate system; For any given triple , the relationship between the head entity h, relation r and tail entity t is shown as follows: ， in, are two weight vectors, are bias vectors that satisfy , represents a set of k-dimensional real numbers; the symbol stands for element-wise product; For each triple , construct the scoring function as shown below: ， in, represents the 1-norm, that is, the sum of the absolute values ​​of the elements of the vector; if If the score of is low, it means that the possibility of the relationship r between the head entity h and the tail entity t is high, otherwise it is low.

3. The method for constructing a multi-label disease data classification model based on representation learning according to claim 2, characterized in that: In the embedding space, the relation r has different connection modes to meet the corresponding requirements, as follows: When the relationship r is a symmetric mode, that is, the relationship between the two entities is mutual, if A has this relationship with B, then B should also have this relationship with A. In this case, ; When the relationship r is an antisymmetric mode, that is, if A and B have this relationship, then B and A do not have this relationship. At this time, it is required that in any dimension i, ; When the relation r is in reverse mode, that is, if there is , then there exists , at this time it is required ; When the relation r is in composite mode, that is, if there is a relation and ,but , at this time it is required .

4. The method for constructing a multi-label disease data classification model based on representation learning according to claim 2, characterized in that: S3 is specifically: S31. Considering the differences in the impact of different entities on the target entity, the attention mechanism is used to calculate the weight scores of different entities, as shown in the following formula: ， in, is an exponential function ; is the temperature parameter, which is used to control the distribution range of attention scores; S32, realize the aggregation of neighbor information centered on the head entity h; S321. For each order of neighbor nodes, a hierarchical attention mechanism is designed to dynamically adjust the weights of neighbor nodes of different orders; ， ， in, is the hierarchical attention score; is the hierarchical weight coefficient; is a learnable parameter used to indicate the importance of the t-th order neighbor node; Indicates when The learnable parameters when ; S322, triples consisting of the head entity h and its first-order neighbors Traverse and fully integrate the embedding information of the head entity h itself and the embedding information of its first-order neighbor nodes to obtain the first-order aggregation vector : ， ， in, is a trainable weight matrix, is the dimension of the input entity embedding vector h, t, is the dimension of the output vector, and Represents the initial embedding vectors of the head entity h and the tail entity t; A set of first-order triples representing the head entity h, represents the first-order neighbor node; The function is used to perform nonlinear transformation on the aggregated vector to avoid gradient disappearance, where x represents the input value of the function; is the cumulative distribution function of the standard normal distribution; S323, adopting a step-by-step iterative method to gradually incorporate higher-order neighbor information, and realize aggregation of high-order neighbor information centered on the head entity h; ； in, is the central node h The order embedding expression, is the central node h The order embedding expression, The central node h is The first-order neighbor node embedding on the .

5. The method for constructing a multi-label disease data classification model based on representation learning according to claim 4, characterized in that: S4 is specifically: After the graph convolution network of the 0th layer is propagated, the disease nodes and patient nodes in the medical knowledge graph are all obtained from the 0th layer to the The embedding vector of the layer is integrated to obtain the complete embedding vector of the patient node p and the disease node d and : ， in, , Represents the patient node p and disease node d in the The embedding vector of the layer; the prediction score between the two is calculated by the dot product operation of the vectors: ， in, is transposed.

6. The method for constructing a multi-label disease data classification model based on representation learning according to claim 5, characterized in that: S5 is as follows: S51. Combined with the self-adversarial negative sampling strategy, dynamic sampling and loss weighting based on node degree are designed to construct the knowledge graph loss function; ， ， in, yes Negative sampling samples, i represents the index of all negative samples, , is the scoring function of negative triples, j represents the index of the current negative sample, is the current negative sample sampling probability, is the temperature parameter; Indicates the degree of the sample node, that is, the number of relationships that connect the node to other nodes; Represents the maximum degree in the knowledge graph; is a hyperparameter used to adjust the influence strength of degree; is a hyperparameter, which is expressed by Dynamically adjust the sampling distribution according to the change of Constructing knowledge graph loss function , to optimize the representation of entities and relations in vector space: ， in, is a fixed interval; is a smoothing factor to prevent the denominator from being 0; n represents the number of negative samples corresponding to each positive sample; is the activation function; S52. Constructing disease data classification loss function : ， in, is a positive sample, is a negative sample, is the sigmoid function; S53. Construct a comprehensive loss function for multi-label disease data classification; as shown in the following formula: ， in, Represents a node collection size; is the regularization coefficient, which is used to control the strength of the regularization term; Represents the L2 norm square of all node embedding vectors; by minimizing the comprehensive loss function The model is trained and optimized to achieve the best accuracy in data recognition and classification.

Images