Atmogram completion method and system based on adjacent node path retrieval

Abstract

The invention relates to the technical field of knowledge maps, in particular to a map completion method and system based on adjacent node path retrieval. The method comprises the following steps: determining a node set with a relationship among nodes through a constructed knowledge graph, forming a positive sample and a negative sample to train a node path number prediction model, and calculating a relationship probability value between a certain node and any other node in the knowledge graph through the node relationship prediction model. And taking the node corresponding to the maximum value of the calculated relation probability values as the relation node of the certain node, obtaining the number of relation paths, and complementing the knowledge graph.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and in particular to a graph completion method and system based on adjacent node path retrieval. Background Art

[0002] With the rapid growth of medical data, how to effectively integrate, manage and utilize this data has become a major challenge for the medical industry. As a technology that combines structured data with unstructured data, knowledge graphs have broad application prospects in the medical field. However, existing medical knowledge graphs often have problems such as incomplete data and missing relationships, which affects their application value in clinical decision-making, drug development and other aspects.

[0003] Big models are used in more and more scenarios due to their excellent performance in various tasks, and have shown great potential in the medical field. However, directly applying big models to knowledge graph completion still faces many challenges, such as how to efficiently extract useful information from massive data and how to accurately complete missing nodes and relationships in the graph. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention aims to provide a graph completion method and system based on adjacent node path retrieval, aiming to automatically complete the missing information in the graph through a large model and adjacent node path retrieval, thereby improving the completeness and accuracy of the medical knowledge graph.

[0005] In order to solve the above problems, the present invention adopts the following technical solutions:

[0006] In one aspect, the present invention provides a graph completion method based on adjacent node path retrieval, comprising:

[0007] According to the constructed knowledge graph, determine the node set that has a relationship with the current entity node, and the node set includes the nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node;

[0008] Combine the current node with the nodes in the node set that have a relationship with it in pairs, and use the number of paths between the nodes in the pairwise combination as labels to form positive samples; combine the current node with the nodes that have no relationship with the current node in pairs to form negative samples;

[0009] Positive and negative samples are used to train the node path number prediction model. A multi-scale feature extraction network is used to extract the multi-scale features of the fusion results of the input data. An adaptive weight allocation network is used to calculate the weights of the multi-scale features. The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention features and calculate the relationship probability value.

[0010] The node relationship prediction model is used to calculate the relationship probability value between a certain node and any other node in the knowledge graph. The node corresponding to the maximum value of the calculated relationship probability value is used as the relationship node of the certain node, and the number of its relationship paths is obtained to complete the knowledge graph.

[0011] As an implementation method, the fusion result h of the input data v Calculate using the following formula:

[0012] h v =f(x v +x co[v] +h ne[v] +x ne[v] )

[0013] in:

[0014] x v =W1*x+b1

[0015] h ne[v] =W2*x ne +b2

[0016] x co[v] =W3*x ne +b3

[0017] x ne[v] =W4*x ne +b4

[0018] W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated;

[0019] W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v;

[0020] W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v;

[0021] W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node;

[0022] Function f is a nonlinear activation function:

[0023]

[0024] Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

[0025] As an implementable method, the final attention feature includes:

[0026] A multi-scale feature extraction network is used to extract multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract multi-scale features of the fusion result of the input data:

[0027] F i =Conv(h i ,W i )

[0028] Among them, F i represents the features extracted using the i-th convolution kernel, h i is the i-th fusion result, W i The weight of the i-th convolution kernel;

[0029] Adopt adaptive weight allocation network to calculate the weights of multi-scale features:

[0030]

[0031] Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature. The score function is calculated using linear transformation or neural network. M is the number of convolution kernels. α i Represents the weight of the i-th feature, which is normalized by the softmax function;

[0032] The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention feature Att(i):

[0033]

[0034] As an implementation method, the relationship probability value l is calculated v Calculate using the following formula:

[0035] l v =tanh(Att(i)).

[0036] On the other hand, the present invention provides a graph completion system based on adjacent node path retrieval, a node set construction module, a sample construction module, a node relationship prediction model construction module and a completion module;

[0037] The node set construction module is used to determine the node set that has a relationship with the current entity node according to the constructed knowledge graph, and the node set includes the nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node;

[0038] The sample construction module is used to combine the current node with nodes in the node set with which it has a relationship in pairs, and use the number of paths between the nodes in the pairwise combination as a label to form a positive sample; and to combine the current node with nodes that have no relationship with the current node in pairs to form a negative sample;

[0039] The node path number prediction model construction module is used to train the node path number prediction model using positive samples and negative samples, extract multi-scale features of the fusion results of the input data using a multi-scale feature extraction network, calculate the weights of the multi-scale features using an adaptive weight allocation network, perform weighted summation of the multi-scale features and the weights of the multi-scale features to obtain the final attention features, and calculate the relationship probability value;

[0040] The completion module is used to calculate the relationship probability value between a certain node and any other node in the knowledge graph through a node relationship prediction model, take the node corresponding to the maximum value of the calculated relationship probability value as the relationship node of the certain node, and obtain its relationship path number to complete the knowledge graph.

[0041] As an implementation method, the fusion result h of the input data v Calculate using the following formula:

[0042] h v =f(x v +x co[v] +h ne[v] +x ne[v] )

[0043] in:

[0044] x v =W1*x+b1

[0045] h ne[v] =W2*x ne +b2

[0046] x co[v] =W3*x ne +b3

[0047] x ne[v] =W4*x ne +b4

[0048] W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated;

[0049] W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v;

[0050] W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v;

[0051] W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node;

[0052] Function f is a nonlinear activation function:

[0053]

[0054] Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

[0055] As an implementable method, the final attention feature includes:

[0056] A multi-scale feature extraction network is used to extract multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract multi-scale features of the fusion result of the input data:

[0057] F i =Conv(h i ,W i )

[0058] Among them, F i represents the features extracted using the i-th convolution kernel, h i is the i-th fusion result, W i The weight of the i-th convolution kernel;

[0059] Adopt adaptive weight allocation network to calculate the weights of multi-scale features:

[0060]

[0061] Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature. The score function is calculated using linear transformation or neural network. M is the number of convolution kernels. α i Represents the weight of the i-th feature, which is normalized by the softmax function;

[0062] The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention feature Att(i):

[0063]

[0064] As an implementation method, the relationship probability value l is calculated v Calculate by the following formula: v =tanh(Att(i)).

[0065] An intelligent device includes: a transmitter, a receiver, a memory and a processor; the memory is used to store computer instructions; the processor is used to run the computer instructions stored in the memory to implement the graph completion method based on adjacent node path retrieval.

[0066] A storage medium includes: a readable storage medium and computer instructions, wherein the computer instructions are stored in the readable storage medium; the computer instructions are used to implement the graph completion method based on adjacent node path retrieval.

[0067] The beneficial effects of the present invention are as follows: the present invention determines the node set of the existence relationship of the node through the constructed knowledge graph, forms positive samples and negative samples to train the node path number prediction model, calculates the relationship probability value between a certain node and any other node in the knowledge graph through the node relationship prediction model, takes the node corresponding to the maximum value of the calculated relationship probability value as the relationship node of the certain node, and obtains its relationship path number to complete the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of a graph completion method based on adjacent node path retrieval of the present invention.

[0069] Figure 2 A schematic diagram of a graph completion system based on adjacent node path retrieval according to the present invention. DETAILED DESCRIPTION

[0070] The present invention is further described in detail below in conjunction with specific embodiments.

[0071] It should be noted that these embodiments are only used to illustrate the present invention rather than to limit the present invention. Simple improvements to the method based on the concept of the present invention all fall within the scope of protection claimed by the present invention.

[0072] See also Figure 1 , which is a graph completion method based on adjacent node path retrieval, including:

[0073] S100. According to the constructed knowledge graph, determine a node set that has a relationship with the current entity node, where the node set includes nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node.

[0074] For example, for the current entity node, named v, first determine all the node sets lis that are related to the current entity node v, and then count the number of relationship paths between the current node v and all the child nodes in the node set lis. The specific calculation process is:

[0075] Assume that the current node is v, and there are n entities a, b, c...n directly connected to v. The number of paths between v and these directly connected entity nodes a, b, c...n is 1, which means that they are directly connected through only one relationship. For nodes a, b, c...n connected to v, we start from these nodes and look for entity nodes directly connected to a, b, c...n. Assume they are a1, b1, c1...n1. There are two relationships connecting these entity nodes to node v, and the corresponding number of paths is 2. After that, we start from a1, b1, c1...n1 and continue to look for their own directly connected entity nodes. We count the number of paths with v one by one and record them until the pointed node has no directly connected entity node.

[0076] In this way, we obtain all the entity nodes that are directly or indirectly connected to the current entity node v, and the number of paths between these entity nodes and v. We put these entity nodes into a set, recorded as the node set lis that has a relationship with the current entity node.

[0077] S200, the current node and the nodes in the node set with which it has a relationship are combined in pairs, and the number of paths between the nodes in the pairwise combination is used as a label to form a positive sample; the current node and the nodes with no relationship with the current node are combined in pairs to form a negative sample.

[0078] The current node v and all the child nodes in the node set lis are combined in pairs as input data x, and the corresponding path relationship number is used as the label. Such data is used as a positive sample. The minimum path relationship number corresponding to the positive sample is 1. Assume that the data volume of the positive sample is s. At the same time, we randomly extract nodes with a data volume of s from the current node v and the nodes that have no relationship with it in the current knowledge graph, and combine these nodes with the current node v in pairs as negative samples. The label of the negative sample is 0, which means that there is no relationship connection between the current node v and the node of the negative sample.

[0079] S300, using positive samples and negative samples to train the node relationship prediction model, using a multi-scale feature extraction network to extract the multi-scale features of the fusion results of the input data, and using an adaptive weight allocation network to calculate the weights of the multi-scale features, and weighted summing the multi-scale features with the weights of the multi-scale features to obtain the final attention features and calculate the relationship probability value.

[0080] If the relationship probability value is predicted to be 0, it means that the model believes that there is no relationship between the two entities. Otherwise, it is believed that there is a relationship between the two entity nodes. The result of the model prediction is the relationship probability value. Finally, the relationship probability value with the highest probability is selected as the model prediction result. According to the prediction result, the number of nodes and paths under the relationship probability value is obtained.

[0081] The fusion result of input data h v Calculate using the following formula:

[0082] h v =f(x v +x co[v] +h ne[v] +x ne[v] )

[0083] in:

[0084] x v =W1*x+b1

[0085] h ne[v] =W2*x ne +b2

[0086] x co[v] =W3*x ne +b3

[0087] x ne[v] =W4*x ne +b4

[0088] W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated;

[0089] W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v;

[0090] W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v;

[0091] W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node;

[0092] Function f is a nonlinear activation function:

[0093]

[0094] Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

[0095] A multi-scale feature extraction network is used to extract the multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract the multi-scale features of the fusion result of the input data. F The formula for i extraction is as follows:

[0096] F i =Conv(h i ,W i )

[0097] h i is the i-th fusion result, and the convolution kernel set is {W1,W2,…,W M}, where M is the number of convolution kernels, W i represents the weight of the i-th convolution kernel, F i Represents the features extracted using the i-th convolution kernel.

[0098] Adopt adaptive weight allocation network to calculate the weights of multi-scale features:

[0099]

[0100] Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature, which is calculated using linear transformation or neural network, α i Represents the weight of the i-th feature, which is normalized by the softmax function;

[0101] The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention features.

[0102] Calculate the relationship probability value l v Calculate using the following formula:

[0103] l v =tanh(Att(i)).

[0104] S400. Calculate the relationship probability value between a certain node and any other node in the knowledge graph through the node relationship prediction model, take the node corresponding to the maximum value of the calculated relationship probability value as the relationship node of the certain node, and obtain its relationship path number to complete the knowledge graph.

[0105] See also Figure 2, is a graph completion system based on adjacent node path retrieval, comprising a node set construction module 100, a sample construction module 200, a node relationship prediction model construction module 300 and a completion module 400;

[0106] The node set construction module 100 is used to determine the node set that has a relationship with the current entity node according to the constructed knowledge graph, and the node set includes the nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node;

[0107] The sample construction module 200 is used to combine the current node with nodes in the node set with which it has a relationship, and use the number of paths between the nodes in the pairwise combination as a label to form a positive sample; and to combine the current node with nodes with which the current node has no relationship, to form a negative sample;

[0108] The node relationship prediction model construction module 300 is used to train the node path number prediction model using positive samples and negative samples, extract the multi-scale features of the fusion results of the input data using a multi-scale feature extraction network, calculate the weights of the multi-scale features using an adaptive weight allocation network, perform weighted summation of the multi-scale features and the weights of the multi-scale features to obtain the final attention features, and calculate the relationship probability value;

[0109] The completion module 400 is used to calculate the relationship probability value between a certain node and any other node in the knowledge graph through a node relationship prediction model, take the node corresponding to the maximum value of the calculated relationship probability value as the relationship node of the certain node, and obtain its relationship path number to complete the knowledge graph.

[0110] The fusion result of input data h v Calculate using the following formula:

[0111] h v =f(x v +x co[v] +h ne[v] +x ne[v] )

[0112] in:

[0113] x v =W1*x+b1

[0114] h ne[v] =W2*x ne +b2

[0115] x co[v] =W3*x ne +b3

[0116] x ne[v] =W4*x ne +b4

[0117] W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated;

[0118] W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v;

[0119] W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v;

[0120] W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node;

[0121] Function f is a nonlinear activation function:

[0122]

[0123] Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

[0124] The final attention features include:

[0125] A multi-scale feature extraction network is used to extract multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract multi-scale features of the fusion result of the input data:

[0126] F i =Conv(h i ,W i )

[0127] Among them, F i represents the features extracted using the i-th convolution kernel, h i is the i-th fusion result, W i The weight of the i-th convolution kernel;

[0128] Adopt adaptive weight allocation network to calculate the weights of multi-scale features:

[0129]

[0130] Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature. The score function is calculated using linear transformation or neural network. M is the number of convolution kernels. αi Represents the weight of the i-th feature, which is normalized by the softmax function;

[0131] The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention feature Att(i):

[0132]

[0133] Calculate the relationship probability value l v Calculate using the following formula:

[0134] l v =tanh(Att(i)).

[0135] A smart device, comprising: a transmitter, a receiver, a memory and a processor;

[0136] The memory is used to store computer instructions; the processor is used to execute the computer instructions stored in the memory to implement the graph completion method based on adjacent node path retrieval of the present invention.

[0137] A storage medium comprises: a readable storage medium and computer instructions, wherein the computer instructions are stored in the readable storage medium; the computer instructions are used to implement the graph completion method based on adjacent node path retrieval of the present invention.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A graph completion method based on adjacent node path retrieval, characterized in that: include: According to the constructed knowledge graph, determine the node set that has a relationship with the current entity node, and the node set includes the nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node; Combine the current node with the nodes in the node set that have a relationship with it in pairs, and use the number of paths between the nodes in the pairwise combination as labels to form positive samples; combine the current node with the nodes that have no relationship with the current node in pairs to form negative samples; Positive and negative samples are used to train the node relationship prediction model. A multi-scale feature extraction network is used to extract the multi-scale features of the fusion results of the input data. An adaptive weight allocation network is used to calculate the weights of the multi-scale features. The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention features and calculate the relationship probability value. The node relationship prediction model is used to calculate the relationship probability value between a certain node and any other node in the knowledge graph. The node corresponding to the maximum value of the calculated relationship probability value is used as the relationship node of the certain node, and the number of its relationship paths is obtained to complete the knowledge graph.

2. The graph completion method based on adjacent node path retrieval according to claim 1 is characterized in that: The fusion result h of the input data v Calculate using the following formula: h v =f(x v +x co[v] +h ne[v] +x ne[v] ) in: x v =W1*x+b1 h ne[v] =W2*x ne +b2 x co[v] =W3*x ne +b3 x ne[v] =W4*x ne +b4 W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated; W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v; W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v; W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node; Function f is a nonlinear activation function: Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

3. The graph completion method based on adjacent node path retrieval according to claim 1 is characterized in that: The final attention features include: A multi-scale feature extraction network is used to extract multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract multi-scale features of the fusion result of the input data: F i =Conv(h i ,W i ) Among them, F i represents the features extracted using the i-th convolution kernel, h i is the i-th fusion result, W i The weight of the i-th convolution kernel; An adaptive weight allocation network is used to calculate the weights of multi-scale features: Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature. The score function is calculated using linear transformation or neural network. M is the number of convolution kernels. α i Represents the weight of the i-th feature, which is normalized by the softmax function; The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention feature Att(i):

4. The graph completion method based on adjacent node path retrieval according to claim 3 is characterized in that: The calculated relationship probability value l v Calculate using the following formula: l v = tanh(At(i)).

5. A graph completion system based on adjacent node path retrieval, characterized in that: It includes a node set building module, a sample building module, a node relationship prediction model building module and a completion module; The node set construction module is used to determine the node set that has a relationship with the current entity node according to the constructed knowledge graph, and the node set includes the nodes that have a relationship with the current entity node and the number of paths between the nodes and the current entity node; The sample construction module is used to combine the current node with nodes in the node set with which it has a relationship in pairs, and use the number of paths between the nodes in the pairwise combination as a label to form a positive sample; and to combine the current node with nodes that have no relationship with the current node in pairs to form a negative sample; The node path number prediction model construction module is used to train the node path number prediction model using positive samples and negative samples, extract multi-scale features of the fusion results of the input data using a multi-scale feature extraction network, calculate the weights of the multi-scale features using an adaptive weight allocation network, perform weighted summation of the multi-scale features and the weights of the multi-scale features to obtain the final attention features, and calculate the relationship probability value; The completion module is used to calculate the relationship probability value between a certain node and any other node in the knowledge graph through a node relationship prediction model, take the node corresponding to the maximum value of the calculated relationship probability value as the relationship node of the certain node, and obtain its relationship path number to complete the knowledge graph.

6. The graph completion system based on adjacent node path retrieval according to claim 5 is characterized in that: The fusion result h of the input data v Calculate using the following formula: h v =f(x v +x co[v] +h ne[v] +x ne[v] ) in: x v =W1*x+b1 h ne[v] =W2*x ne +b2 x co[v] =W3*x ne +b3 x ne[v] =W4*x ne +b4 W1 is the weight parameter matrix, b1 is the bias parameter matrix, x v The three-dimensional matrix representation of the current node x calculated; W2 is the weight parameter matrix, b2 is the bias parameter matrix, h ne[v] It is the three-dimensional matrix representation of the adjacent nodes of the current node v; W3 is the weight parameter matrix, b3 is the bias parameter matrix, x co[v] It is the three-dimensional matrix representation of the edges of the adjacent nodes of the current node v; W4 is the weight parameter matrix, b4 is the bias parameter matrix, x ne[v] is a three-dimensional matrix representation of the features of the neighboring nodes of the current node v; ne is the adjacent node of the current node; Function f is a nonlinear activation function: Among them, m is the input of the current function, and exp is the exponential function with the natural constant e as the base.

7. The graph completion system based on adjacent node path retrieval according to claim 5 is characterized in that: The final attention features include: A multi-scale feature extraction network is used to extract multi-scale features of the fusion result of the input data, wherein the multi-scale feature extraction network uses multiple convolution kernels to extract multi-scale features of the fusion result of the input data: F i =Conv(h i ,W i ) Among them, F i represents the features extracted using the i-th convolution kernel, h i is the i-th fusion result, W i The weight of the i-th convolution kernel; An adaptive weight allocation network is used to calculate the weights of multi-scale features: Among them, s(F i ) represents the score function of the i-th feature, s(F j ) represents the score function of the jth feature. The score function is calculated using linear transformation or neural network. M is the number of convolution kernels. α i Represents the weight of the i-th feature, which is normalized by the softmax function; The multi-scale features are weighted and summed with the weights of the multi-scale features to obtain the final attention feature Att(i):

8. The graph completion system based on adjacent node path retrieval according to claim 5 is characterized in that: The calculated relationship probability value l v Calculate using the following formula: l v = tanh(At(i)).

9. A smart device, characterized in that: include: transmitter, receiver, memory and processor; The memory is used to store computer instructions; The processor is used to execute the computer instructions stored in the memory to implement the graph completion method based on adjacent node path retrieval as described in any one of claims 1 to 4.

10. A storage medium, characterized in that: include: A readable storage medium and computer instructions, wherein the computer instructions are stored in the readable storage medium; The computer instructions are used to implement the graph completion method based on adjacent node path retrieval as described in any one of claims 1 to 4.

Images