Network representation learning across medical data sources

Abstract

The present disclosure proposes a network representation learning method across medical data sources, comprising: S1, generating medical network data comprising a source network and a target network; S2, randomly sampling a set number of nodes from the source network and the target network; S3, obtaining an L-layer neural network, and calculating the structural features and expression features of the source network and the target network respectively for each layer, and calculating the distance loss between the network features of the source network and the target network; S4, obtaining the output of the source network in the L-layer neural network, and calculating the loss value according to the classification loss and the distance loss, and updating the parameters of the algorithm according to the back propagation algorithm; S5, repeating steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification no longer rises within multiple iterations. The present disclosure considers the problem of inconsistent data distribution between different hospital data sources, and through the extraction of network structural information and node attribute information and the minimization of feature distance, the information loss is compensated, which has a wide application space.

Description

Technical Field

[0001] This disclosure belongs to the field of medical technology, and in particular relates to a network representation learning method across medical data sources. Background Technology

[0002] As deep learning technology matures, its application in medical settings is becoming increasingly possible. By modeling patient cases and utilizing deep learning, doctors can infer patient conditions, thus advancing the medical field. In medical settings, our research subjects often include patients, symptoms, diseases, and medications, which have complex logical relationships. These relationships are difficult to characterize directly using a simple deep fully connected network. Furthermore, considering that network embedding representation learning algorithms can intuitively represent the connections between objects and possess powerful reasoning capabilities, we can leverage this technique to model and solve problems from a network perspective. Specifically, we can view each involved object as a node in a network; edges connecting nodes indicate a relationship. For example, the relationship between nodes could be a treatment relationship between a disease and medication, or a manifestation of a disease and its symptoms. By abstracting complex medical relationships into a network and then extracting information from it using network embedding representation algorithms, we can achieve the goal of inferring a patient's condition.

[0003] Existing network embedding representation algorithms can be mainly divided into two categories: First, transductive representation learning algorithms. Given a target network, transductive representation learning algorithms directly optimize the representation vector of each node through node attributes and network relationships, such as DeepWalk and Node2vec. Second, inductive representation learning algorithms. Inductive representation learning algorithms often learn a mapping function. Given the attributes of the input node and its neighbors, the representation vector of the node can be inferred through the mapping function, such as GCN, GraphSAGE, and GAT.

[0004] In real-world medical scenarios, medical data often comes from different hospitals. This leads to inconsistent distributions of patient data across different hospitals. For example, the same illness like the common cold might be a cold caused by exposure to cold in the south, while in the north it might be a heat-related cold caused by indoor heating. Furthermore, the same disease might be treated with multiple medications for the same pathology, and different hospitals may have different prescribing habits. These realities result in different data distributions across hospitals. Therefore, when aiming to utilize diverse data to aid model learning, directly using data from multiple different medical data sources without considering their inconsistent distribution can potentially lead to biased knowledge.

[0005] However, existing algorithms are not well-suited for solving this type of network representation learning problem across medical data sources. Specifically:

[0006] (1) For the transductive algorithm, since the transductive algorithm directly optimizes the node representation vectors in a network composed of data from a certain hospital, it cannot directly infer the node representation vectors in a new network composed of data from another hospital. Therefore, the transductive algorithm has no available knowledge that can be used for network learning across medical data sources.

[0007] (2) For inductive algorithms, although they consider learning a mapping function between node attributes and structural information when modeling, which can naturally perform network inference across medical data sources, inductive algorithms do not take into account that the data distribution between networks is different. The patterns or knowledge inferred from the medical network of one hospital may not be well applied to the medical network of another hospital. Therefore, inductive algorithms also have certain defects in the problem of learning network representations across medical data sources.

[0008] Therefore, existing technologies need to be improved.

[0009] The above background information is provided only to help understand this disclosure and does not imply an acknowledgment or endorsement that any content mentioned is part of the general knowledge relative to this disclosure. Summary of the Invention

[0010] To address the aforementioned technical issues, this disclosure proposes a network representation learning method across medical data sources.

[0011] Based on one aspect of the embodiments of this disclosure, a network representation learning method across medical data sources is disclosed, comprising:

[0012] S1, Generate medical network data including a source network and a target network. The source network is generated from the medical records of a certain hospital, and the target network is generated from the medical records of another hospital different from the source network. The medical network data includes the patient's medical record information and constructs network relationships between symptoms, diseases, drugs, and diagnostic methods.

[0013] S2, randomly sample a set number of nodes from the source network and the target network respectively, the number of nodes sampled being related to the degree of the medical network;

[0014] S3, obtain an L-layer neural network from step S2, and calculate the structural and representational features of the source and target networks for each layer, and calculate the distance loss between the network features of the source and target networks;

[0015] S4: Obtain the output of the source network in the L-layer neural network from S3, calculate the loss value based on the classification loss and distance loss, and update the algorithm parameters according to the backpropagation algorithm;

[0016] S5. Repeat steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification no longer increases within multiple iterations.

[0017] In another embodiment of the network representation learning method across medical data sources based on this disclosure, step S1, generating medical network data including a source network and a target network, includes:

[0018] Retrieve the files containing the nodes and edges of the source network, and the files containing the nodes and edges of the target network;

[0019] Based on the files storing the nodes and edges of the source network and the files storing the nodes and edges of the target network, the medical network data including the source network and the target network is generated.

[0020] In another embodiment of the network representation learning method based on the present disclosure across medical data sources, step S3 involves obtaining an L-layer neural network from step S2, and calculating the structural and representational features of the source and target networks for each layer. The calculation of the distance loss between the network features of the source and target networks includes:

[0021] S30, input the node features of the source network and the target network into the L-layer neural network;

[0022] S31, In any layer l of the L-layer neural network, the node expression feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are processed by a message aggregation module to obtain the new expression feature vector of the current node.

[0023] S32, through the network alignment module across medical data sources, calculates the distance loss value between node features from the source network and the target network in the current layer;

[0024] S33. Repeat steps S31 to S32 L times to obtain the final node feature vectors of the source network and the target network, as well as the accumulated structural feature distance loss and representation feature distance loss of L layers.

[0025] In another embodiment of the network representation learning method based on the present disclosure across medical data sources, step S31, in any layer of the L-layer neural network... l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are then processed by a message aggregation module to obtain a new representation feature vector for the current node, including:

[0026] Obtain any layer of an L-layer neural network l The feature vector of any node u, and the feature vectors of the neighboring nodes of node u;

[0027] Update the expression feature vector of node u using the expression feature vectors of the neighboring nodes;

[0028] The updated feature vector of node u is propagated to the feature vectors of node u's neighboring nodes to obtain the updated feature vectors of node u's neighboring nodes.

[0029] In another embodiment of the network representation learning method based on the present disclosure across medical data sources, step S31, in any layer l of the L-layer neural network, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are processed by a message aggregation module to obtain a new representation feature vector for the current node, including:

[0030] The node representation feature vector of each network is processed by a message routing module to obtain the structural feature representation as follows:

[0031] ;

[0032] ;

[0033] In the formula, For nodes In an L-layer neural network, the first Layer structural feature vectors In an L-layer neural network, the first... The source and target network feature vectors of the layers; the feature vector of layer 0 is derived from the original feature vectors of the nodes. express, For the first The parameter matrix involved in the message routing module of the layer, For the first The parameter matrix involved in the message routing module of the layer, For activation function, This is the direct concatenation operation between two vectors. For nodes The set of directly connected neighbors. For nodes Transmit to node Message weight;

[0034] The structural features, after being processed by a message aggregation module, yield a new expression feature vector for the current node, represented as follows:

[0035] ;

[0036] ;

[0037] In the formula, and This is the parameter matrix involved in the message aggregation module. This represents the vector at the node aggregation level.

[0038] In another embodiment of the network representation learning method across medical data sources based on this disclosure, step S32, calculating the distance loss value between node features from the source network and the target network in the current layer through the network alignment module across medical data sources, includes:

[0039] The structural feature distance loss for each layer is:

[0040] ;

[0041] In the formula, The structural feature vectors of the source network and the target network and The distribution, It is a distance function used to calculate the structural feature vector. and The expected distance;

[0042] The feature distance loss for each layer is:

[0043] ;

[0044] In the formula, The feature vectors representing the nodes of the source and target networks. and The distribution, It is a distance function used to calculate the feature vector representing a node. and The expected distance.

[0045] In another embodiment of the network representation learning method based on the cross-medical data source of this disclosure, step S33, repeating steps S31 to S32 L times, to obtain the final node feature vectors of the source network and the target network, and the accumulated structural feature distance loss and representation feature distance loss of L layers, includes:

[0046] The cumulative structural feature distance loss of layer L is:

[0047] ;

[0048] The cumulative feature distance loss of layer L is:

[0049] .

[0050] Compared with the prior art, this disclosure has the following advantages:

[0051] This paper adopts a network representation learning method across medical data sources, which takes into account the problem of inconsistent data distribution between different hospital data sources. It compensates for the information loss caused by inconsistency by minimizing feature distance, and uses classification loss to ensure that the vector representation of the calculated nodes can accurately label the nodes representing diseases and symptoms. This allows for full utilization of multi-source medical data and has broad application prospects in the medical field. Attached Figure Description

[0052] Figure 1 A flowchart illustrating an embodiment of the network representation learning method across medical data sources proposed in this disclosure;

[0053] Figure 2 This is a flowchart of another embodiment of the network representation learning method across medical data sources proposed in this disclosure. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0055] The following description, in conjunction with the accompanying drawings and embodiments, provides a more detailed explanation of a network representation learning method across medical data sources provided in this disclosure.

[0056] The network representation learning method across medical data sources proposed in this application can be applied to the following scenarios.

[0057] There are two hospitals, A and B, which are not affiliated with the same institution. Both hospitals want to construct their own disease symptom networks. Hospital A's disease symptom network labels the name of the disease or symptom corresponding to each node. However, due to time constraints, Hospital B only constructed the network structure for the disease symptoms without labeling the name of the disease or symptom for each node. Therefore, Hospital B wants to use Hospital A's disease symptom network to train vector representations of the nodes in its own disease symptom network, and then infer the names of the diseases or symptoms represented by each node in its network. Specifically, Hospital B provides its disease symptom network to Hospital A, and Hospital A needs to use its own disease symptom network and the received network from Hospital B to train vector representations of the nodes in Hospital B's disease symptom network. We can call Hospital A's disease symptom network the source network and Hospital B's network the target network.

[0058] Existing network embedding representation methods cannot solve this problem. Existing network embedding representation algorithms can be mainly divided into two categories: transductive representation learning algorithms and inductive representation learning algorithms. Transductive representation learning algorithms require knowledge of what disease or symptom each node in the disease symptom network of Hospital B represents, therefore they cannot be used in this task. Inductive representation learning algorithms learn a mapping function; given the attributes of an input node and its neighbors, the representation vector of the node can be inferred from the mapping function. However, it cannot perceive the inconsistency in data distribution between the source and target networks of the disease symptoms, so the mapping function inferred from the source network may not be well applied to the target network.

[0059] Based on the technical problems existing in the above-mentioned application fields, embodiments of this application disclose a network representation learning method across medical data sources.

[0060] Figure 1 A flowchart of an embodiment of the network representation learning method across medical data sources proposed in this disclosure is shown below. Figure 1 As shown, the network representation learning method across medical data sources is as follows:

[0061] S1, Generate medical network data including a source network and a target network. The source network is generated from the medical records of a certain hospital, and the target network is generated from the medical records of another hospital different from the source network. The medical network data includes the patient's medical record information and constructs network relationships between symptoms, diseases, drugs, and diagnostic methods.

[0062] Specifically, in one embodiment of this application, step S1, generating medical network data including a source network and a target network, includes:

[0063] Obtain the files storing the nodes and edges of the source network, and the files storing the nodes and edges of the target network; specifically, the files storing the nodes and edges of the source network are provided by the hospital in the source network, where the name of the disease or symptom corresponding to each node is known; the files storing the nodes and edges of the target network are provided by the hospital in the target network, where the disease or symptom corresponding to each node is unknown.

[0064] Based on the files storing the nodes and edges of the source network and the files storing the nodes and edges of the target network, the medical network data including the source network and the target network is generated. Specifically, the correspondence between nodes and diseases or symptoms is read from the medical network data of the source network to obtain the name of the disease or symptom corresponding to each node in the source network, and the disease symptoms are read from the medical network data of the target network.

[0065] S2, randomly sample a set number of nodes from the source network and the target network respectively. The number of nodes sampled is related to the degree of the medical network. Specifically, taking disease diagnosis as an example, it is necessary to iterate L times from each disease node to collect its neighboring nodes, and use the node attributes corresponding to the collected nodes as the input data of the algorithm.

[0066] S3. Obtain an L-layer neural network from step S2, and calculate the structural and representational features of the source and target networks for each layer, and calculate the distance loss between the network features of the source and target networks.

[0067] Specifically, for an L-layer neural network, when performing any layer... l During propagation, for any node u in the source network and the target network, the message routing module can be used to obtain the structural and representational features of the current propagating node u. Based on the structural features of node u and its neighboring nodes, the message weights of node u when receiving information from each of its neighbors can be calculated through the attention mechanism.

[0068] Specifically, Figure 2 A flowchart of another embodiment of the network representation learning method across medical data sources proposed in this disclosure, as follows: Figure 2 As shown, step S3 obtains an L-layer neural network from step S2, and calculates the structural and representational features of the source and target networks for each layer, and calculates the distance loss between the network features of the source and target networks, including:

[0069] S30, input the node features of the source network and the target network into the L-layer neural network;

[0070] S31, in any layer of the L-layer neural network l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are processed by a message aggregation module to obtain a new representation feature vector for the current node.

[0071] Specifically, in one embodiment of this application, step S31 occurs in any layer of an L-layer neural network. l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are then processed by a message aggregation module to obtain a new representation feature vector for the current node, including:

[0072] Obtain any layer of an L-layer neural network l The feature vector of any node u, and the feature vectors of the neighboring nodes of node u;

[0073] Update the expression feature vector of node u using the expression feature vectors of the neighboring nodes;

[0074] The updated feature vector of node u is propagated to the feature vectors of node u's neighboring nodes to obtain the updated feature vectors of node u's neighboring nodes.

[0075] Specifically, in one embodiment of this application, step S31 is performed in any layer of the L-layer neural network. l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are then processed by a message aggregation module to obtain a new representation feature vector for the current node, including:

[0076] The node representation feature vector of each network is processed by a message routing module to obtain the structural feature representation as follows:

[0077] ;

[0078] ;

[0079] In the formula, For nodes In an L-layer neural network, the first Layer structural feature vectors In an L-layer neural network, the first... The source and target network feature vectors of the layers; the feature vector of layer 0 is derived from the original feature vectors of the nodes. express, For the first The parameter matrix involved in the message routing module of the layer, For the first The parameter matrix involved in the message routing module of the layer, For activation function, This is the direct concatenation operation between two vectors. For nodes The set of directly connected neighbors. For nodes Transmit to node Message weight;

[0080] The structural features, after being processed by a message aggregation module, yield a new representation of the current node's feature vector, which is as follows:

[0081] ;

[0082] ;

[0083] In the formula, and This is the parameter matrix involved in the message aggregation module. This represents the vector at the node aggregation level.

[0084] S32, through the network alignment module across medical data sources, calculates the distance loss value between node features from the source network and the target network in the current layer;

[0085] Specifically, the distance loss is calculated through the network alignment module. This distance loss comprises two aspects: structural feature distance loss and representational feature distance loss. During the propagation of the L-layer neural network, the structural feature distance loss and representational feature distance loss between the source and target networks are calculated after each propagation. These two losses measure the discrepancy between the distributions of structural and representational features of nodes representing diseases and symptoms in the source and target networks, respectively. By minimizing the distance between the feature distributions of these two networks, the information loss caused by inconsistent data distributions between different hospital data sources is compensated for. Finally, the structural feature distance loss and representational feature distance loss after each propagation of the L-layer neural network are summed to obtain the total distance loss.

[0086] Specifically, in one embodiment of this application, step S32, which calculates the distance loss value between node features from the source network and the target network in the current layer through a network alignment module across medical data sources, includes:

[0087] The structural feature distance loss for each layer is:

[0088] ;

[0089] In the formula, The structural feature vectors of the source network and the target network and The distribution, It is a distance function used to calculate the structural feature vector. and The expected distance; this formula represents the gap between the distribution of structural features of nodes representing diseases and symptoms in the source network and the target network.

[0090] The feature distance loss for each layer is:

[0091] ;

[0092] In the formula, The feature vectors representing the nodes of the source and target networks. and The distribution, It is a distance function used to calculate the feature vector representing a node. and The expected distance is expressed as the difference between the distribution of the expression features of nodes representing diseases and symptoms in the source and target networks.

[0093] S33. Repeat steps S31 to S32 L times to obtain the final node feature vectors of the source network and the target network, as well as the accumulated structural feature distance loss and representation feature distance loss of L layers.

[0094] Specifically, during the propagation process of the L-layer neural network, after each propagation, the structural feature distance loss and the expression feature distance loss are calculated once based on the structural feature vector and the expression feature vector of the node representing the disease or symptom at that time.

[0095] Specifically, in one embodiment of this application, step S33, repeating steps S31 to S32 L times to obtain the final node feature vectors of the source network and the target network, and the accumulated structural feature distance loss and representation feature distance loss of L layers, includes:

[0096] The cumulative structural feature distance loss of layer L is:

[0097] ;

[0098] The cumulative feature distance loss of layer L is:

[0099] .

[0100] S4 obtains the output of the source network in the L-layer neural network from S3, calculates the loss value based on the classification loss and distance loss, and updates the algorithm parameters according to the backpropagation algorithm.

[0101] Specifically, based on the calculated total loss value, the backpropagation algorithm is used to update the various parameters in the algorithm, and then the next iteration begins.

[0102] S5. Repeat steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification no longer increases within multiple iterations.

[0103] Specifically, each iteration propagates through L layers. This process is repeated until the entire algorithm converges, ensuring the loss value no longer decreases over multiple iterations, thus obtaining vector representations of each node in the source and target networks for the disease symptoms. Using these vector representations of each node in the target network, the disease or symptom name represented by each node is inferred, thereby completing the node labeling task in the target network for the disease symptoms.

[0104] It will be apparent to those skilled in the art that the embodiments of this disclosure are not limited to the details of the exemplary embodiments described above, and that the embodiments of this disclosure can be implemented in other specific forms without departing from the spirit or essential characteristics of the embodiments of this disclosure. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the embodiments of this disclosure is defined by the appended claims rather than the foregoing description. Therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be encompassed within the embodiments of this disclosure. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units, modules, or devices recited in the system, apparatus, or terminal claims may also be implemented by the same unit, module, or device through software or hardware. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure and are not intended to limit the scope thereof. Although the present disclosure has been described in detail with reference to the above preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present disclosure should not depart from the spirit and scope of the technical solutions of the present disclosure.

Claims

1. A network representation learning method across medical data sources, characterized in that, include: S1, Generate medical network data including a source network and a target network. The source network is generated from medical records of a certain hospital, and the name of the disease or symptom corresponding to each node in the source network is known. The target network is generated from medical records of another hospital different from the source network, and the disease or symptom corresponding to each node in the target network is unknown. The medical network data includes patient medical record information, and a network relationship is constructed between symptoms, diseases, drugs, and diagnostic methods. The generation of the medical network data including the source network and the target network includes: obtaining files storing the nodes and edges of the source network and files storing the nodes and edges of the target network; and generating the medical network data including the source network and the target network based on the files storing the nodes and edges of the source network and the target network. S2, randomly sample a set number of nodes from the source network and the target network respectively, the number of nodes sampled being related to the degree of the medical network; S3, obtain an L-layer neural network from step S2, and perform the following for each layer of the L-layer neural network: input the node representation feature vectors of the source network and the target network into the message routing module, and the message routing module outputs the structural features of the node; input the structural features into the message aggregation module, and the message aggregation module outputs the new representation feature vector of the node; calculate the distance loss between the network features of the source network and the target network; wherein, the node representation feature vector of the 0th layer is represented by the original feature vector of the node; S4: Obtain the node representation feature vectors of each node in the source network output by the L-layer neural network from S3, calculate the loss value based on the classification loss and distance loss, and update the algorithm parameters according to the backpropagation algorithm; the classification loss is determined based on the disease or symptom classification results of the nodes in the source network. S5. Repeat steps S2-S4 until the entire algorithm converges, so that the accuracy of the algorithm for disease classification no longer increases within multiple iterations.

2. The network representation learning method across medical data sources according to claim 1, characterized in that, Step S3 involves obtaining an L-layer neural network from step S2, and calculating the structural and representational features of the source and target networks for each layer. The distance loss between the network features of the source and target networks is calculated, including: S30, input the node features of the source network and the target network into the L-layer neural network; S31, in any layer of the L-layer neural network l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are processed by a message aggregation module to obtain a new representation feature vector for the current node. S32, through the network alignment module across medical data sources, calculates the distance loss value between node features from the source network and the target network in the current layer; S33. Repeat steps S31 to S32 L times to obtain the final node feature vectors of the source network and the target network, as well as the accumulated structural feature distance loss and representation feature distance loss of L layers.

3. The network representation learning method across medical data sources according to claim 2, characterized in that, S31, in any layer of the L-layer neural network l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are then processed by a message aggregation module to obtain a new representation feature vector for the current node, including: Obtain any layer of an L-layer neural network l The feature vector of any node u, and the feature vectors of the neighboring nodes of node u; Update the expression feature vector of node u using the expression feature vectors of the neighboring nodes; The updated feature vector of node u is propagated to the feature vectors of node u's neighboring nodes to obtain the updated feature vectors of node u's neighboring nodes.

4. The network representation learning method across medical data sources according to claim 2, characterized in that, In step S31, at any layer of the L-layer neural network l In this process, the node representation feature vector of each network is processed by a message routing module to obtain structural features, and the structural features are then processed by a message aggregation module to obtain a new representation feature vector for the current node, including: The node representation feature vector of each network is processed by a message routing module to obtain the structural feature representation as follows: In the formula, For nodes In an L-layer neural network, the first Layer structural feature vectors In an L-layer neural network, the first... The source and target network feature vectors of the layers; the feature vector of layer 0 is derived from the original feature vectors of the nodes. express, and It is a parameter matrix involved in the message routing module. For activation function, This is the direct concatenation operation between two vectors. For nodes The set of directly connected neighbors. For nodes Transmit to node Message weight; The structural features, after being processed by a message aggregation module, yield a new expression feature vector for the current node, represented as follows: ； ； In the formula, and This is the parameter matrix involved in the message aggregation module. This represents the vector at the node aggregation level.

5. The network representation learning method across medical data sources according to claim 4, characterized in that, Step S32, which calculates the distance loss value between node features from the source network and the target network in the current layer through the cross-medical data source network alignment module, includes: The structural feature distance loss for each layer is: ； In the formula, The structural feature vectors of the source network and the target network and The distribution, It is a distance function used to calculate the structural feature vector. and The expected distance; The feature distance loss for each layer is: ； In the formula, The feature vectors representing the nodes of the source and target networks. and The distribution, It is a distance function used to calculate the feature vector representing a node. and The expected distance.

6. The network representation learning method across medical data sources according to claim 5, characterized in that, Step S33 repeats steps S31 to S32 L times to obtain the final node feature vectors of the source network and the target network, as well as the accumulated structural feature distance loss and representation feature distance loss of L layers, including: The cumulative structural feature distance loss of layer L is: ； The cumulative feature distance loss of layer L is: 。

Images