Link prediction method and device and storage medium

Abstract

The invention discloses a link prediction method and device and a storage medium, and the method comprises the steps: obtaining a preset first data set, and determining N nodes and known edge connection information between the nodes according to the preset first data set, N being a positive integer greater than or equal to 3; determining initial feature vectors of the N nodes; hyperedges between the nodes are generated according to the initial feature vectors of the N nodes and known edge connection information between the nodes; constructing a hypergraph according to the N nodes and the generated hyperedges, and performing hypergraph representation learning on the hypergraph; and performing link prediction between the nodes according to the hypergraph representation learning result.

Description

Technical Field

[0001] This article relates to the intersection of computer and medicine, and in particular to a link prediction method, device and storage medium. Background Art

[0002] Link prediction in graph data mining is used to predict whether there is a link between two nodes in the graph. Known link prediction methods generally include methods based on similarity, methods based on maximum likelihood, methods based on machine learning, and methods based on graph neural networks. Among them, the method based on graph neural networks is to construct known nodes and links into a normal graph, and then learn the representation vector of the nodes in the graph through the graph neural network, and perform link prediction based on the learned node representation vector. Or, the known information is constructed into a hypergraph based on K nearest neighbor equidistance functions, and then the node representation vector is learned through the hypergraph neural network for link prediction. Summary of the invention

[0003] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0004] The embodiments of the present application provide a link prediction method, device and storage medium. The link prediction method can not only reflect the high-order relationship of node feature information, but also reflect the high-order relationship of network structure information, which helps to improve the accuracy of link prediction.

[0005] A link prediction method provided in an embodiment of the present application includes:

[0006] Obtain a preset first data set, and determine N nodes and known edge information between the nodes according to the preset first data set, where N is a positive integer greater than or equal to 3;

[0007] Determine the initial feature vectors of the N nodes;

[0008] Generate hyperedges between nodes according to the initial feature vectors of the N nodes and the known edge information between the nodes;

[0009] Constructing a hypergraph according to the N nodes and the generated hyperedges and performing hypergraph representation learning on the hypergraph;

[0010] Link prediction between nodes is performed based on the hypergraph representation learning result.

[0011] In an exemplary embodiment, generating a hyperedge between nodes according to known edge information between the nodes includes:

[0012] Constructing a common graph according to the known edge information between the nodes;

[0013] Determining an adjacency matrix of the general graph;

[0014] Establishing a linear regression model for each node according to the adjacency matrix;

[0015] Solving the optimization target of the linear regression model of each node;

[0016] Generate hyperedges based on the solution results.

[0017] In an exemplary embodiment, the linear regression model is:

[0018] B(i)=X(i)Z(i)+b(i);

[0019] Wherein, B represents the adjacency matrix, B(i) is the i-th row of the adjacency matrix, representing the structural information of node i, X(i) represents the structural information of other nodes except node i, Z(i) is the regression coefficient, representing the correlation coefficient between node i and other nodes, and b(i) is the random error term, 1≤i≤N.

[0020] In an exemplary embodiment, the optimization target is determined according to the following formula:

[0021] min Z(i) ||B(i)-X(i)Z(i)||+λ||Z(i)||;

[0022] Among them, ||·|| represents the L1 norm, min represents the minimum value, λ is the regularization parameter, 0<λ<1.

[0023] In an exemplary embodiment, generating a hyperedge according to the solution result includes:

[0024] Generate a hyperedge between the node corresponding to the non-zero element in Z(i) and node i.

[0025] In an exemplary embodiment, performing hypergraph representation learning on the hypergraph includes:

[0026] Using a hypergraph constructed by the N nodes and the generated hyperedges as an initial hypergraph;

[0027] Performing hypergraph representation learning and structural adjustment on the initial hypergraph through a multi-layer hypergraph convolutional neural network;

[0028] Wherein, each layer of the multi-layer hypergraph convolutional neural network is composed of a layer of hypergraph convolutional neural network.

[0029] In an exemplary embodiment, the performing hypergraph representation learning and structural adjustment on the hypergraph by using a multi-layer hypergraph convolutional neural network includes:

[0030] Input the initial hypergraph into the first layer of the multi-layer hypergraph convolutional neural network to obtain the feature vectors of the nodes in the first layer;

[0031] The input of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the (I + 1)-th layer hypergraph generated according to the feature vectors of the nodes in the I-th layer; the output of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the feature vectors of the nodes in the (I + 1)-th layer, where 1 ≤ I < M and M is the total number of layers of the multi-layer hypergraph convolutional neural network.

[0032] In an exemplary embodiment, generating the (I + 1)-th layer hypergraph according to the feature vectors of the nodes in the I-th layer includes:

[0033] Generating the (I + 1)-th layer hypergraph based on the K-nearest neighbor algorithm according to the feature vectors of the nodes in the I-th layer.

[0034] In an exemplary embodiment, the hypergraph representation learning result is the feature vectors of the nodes in the M-th layer.

[0035] In an exemplary embodiment, performing link prediction between nodes according to the hypergraph representation learning includes:

[0036] Calculating the cosine similarity between the feature vectors of the node pairs to be predicted according to the hypergraph representation learning, and determining whether there is an association between the node pairs to be predicted according to a preset cosine similarity threshold;

[0037] Or,

[0038] Calculating the inner product between the feature vectors of the node pairs to be predicted according to the hypergraph representation learning, and determining whether there is an association between the node pairs to be predicted according to a preset inner product threshold;

[0039] Or,

[0040] Concatenating the feature vectors of the node pairs to be predicted according to the hypergraph representation learning, and inputting the concatenated feature vectors into a binary classifier to determine whether there is an association between the node pairs to be predicted.

[0041] In an exemplary embodiment, the N nodes include miRNA nodes and disease nodes;

[0042] Determining the initial feature vectors of the N nodes includes:

[0043] Determining a disease semantic similarity matrix and a miRNA function similarity matrix;

[0044] Constructing a miRNA-disease adjacency matrix, and respectively determining a disease Gaussian spectral kernel similarity matrix and a miRNA Gaussian spectral kernel similarity matrix according to the miRNA-disease adjacency matrix;

[0045] Integrating the disease semantic similarity matrix and the disease Gaussian spectrum kernel similarity matrix to obtain a disease integrated similarity matrix; integrating the miRNA functional similarity matrix and the miRNA Gaussian spectrum kernel similarity matrix to obtain a miRNA integrated similarity matrix;

[0046] The initial feature vectors of the N nodes are determined according to the disease integrated similarity matrix and the miRNA integrated similarity matrix.

[0047] The embodiment of the present application also provides a link prediction device, including a memory and a processor.

[0048] The memory is used to store a program for link prediction;

[0049] The processor is used to read the program for executing the link prediction and execute the link prediction method described in any one of the above embodiments.

[0050] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to enable the computer to execute the link prediction method described in any of the above embodiments.

[0051] Other aspects will be apparent upon reading and understanding the drawings and detailed description.

[0052] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by implementing the present application. Other advantages of the present application can be realized and obtained by the schemes described in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide an understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.

[0054] Figure 1 A schematic diagram of a link prediction method according to an embodiment of the present application;

[0055] Figure 2 A schematic diagram of a hypergraph constructed for an embodiment of the present application;

[0056] Figure 3 A schematic diagram of performing hypergraph representation learning and structural adjustment on a hypergraph in an embodiment of the present application;

[0057] Figure 4 A schematic diagram of a miRNA-disease association prediction method according to an embodiment of the present application;

[0058] Figure 5A schematic diagram of a link prediction device according to an embodiment of the present application. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the embodiments of the present disclosure will be described in detail in conjunction with the accompanying drawings hereinafter. Note that the embodiments can be implemented in a plurality of different forms. A person of ordinary skill in the art can easily understand the fact that the methods and contents can be transformed into various forms without departing from the purpose and scope of the present disclosure. Therefore, the present disclosure should not be interpreted as being limited to the contents described in the following embodiments. In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments can be arbitrarily combined with each other. In order to keep the following description of the embodiments of the present disclosure clear and concise, the present disclosure omits the detailed description of some known functions and known components. The drawings of the embodiments of the present disclosure only relate to the structures involved in the embodiments of the present disclosure, and other structures can refer to the usual design.

[0060] In the present specification, ordinal numbers such as “first”, “second” and “third” are provided to avoid confusion among constituent elements, and are not intended to limit the number.

[0061] In this specification, for the sake of convenience, the words and phrases indicating the orientation or positional relationship, such as "middle", "upper", "lower", "front", "back", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., are used to illustrate the positional relationship of the constituent elements with reference to the drawings. This is only for the convenience of describing this specification and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as a limitation of the present disclosure. The positional relationship of the constituent elements is appropriately changed according to the direction in which each constituent element is described. Therefore, it is not limited to the words and phrases described in the specification, and can be appropriately replaced according to the situation.

[0062] At present, most of the link prediction methods mainly use the pairwise low-order relationships between nodes in the network, and it is difficult to model the high-order relationships between nodes. Although some methods use the K-nearest neighbor algorithm to cluster node feature information to achieve hypergraph construction, the constructed hypergraph lacks high-order information of the network structure.

[0063] In order to solve the problem that the constructed hypergraph lacks high-order information of the network structure, this application proposes a link prediction method.

[0064] Figure 1 FIG. 1 is a schematic diagram of a link prediction method according to an embodiment of the present application. Figure 1 As shown, the link prediction method includes the following steps 11 to 15:

[0065] Step 11, obtaining a preset first data set, and determining N nodes and known edge information between the nodes according to the preset first data set, where N is a positive integer greater than or equal to 3;

[0066] Step 12, determining the initial feature vectors of the N nodes;

[0067] Step 13, generating hyperedges between nodes according to the initial feature vectors of the N nodes and the known edge information between the nodes;

[0068] Step 14, constructing a hypergraph according to the N nodes and the generated hyperedges and performing hypergraph representation learning on the hypergraph;

[0069] Step 15: Predict links between nodes based on the hypergraph representation learning result.

[0070] The embodiment of the present application not only generates hyperedges based on the initial feature vectors of the nodes, but also generates hyperedges based on the edge information between the nodes, and constructs a hypergraph based on the nodes and the generated hyperedges, so that the constructed hypergraph not only reflects the high-order relationship between the node attributes, but also can reflect the high-order relationship of the known network structure. Link prediction is performed based on the constructed hypergraph, which helps to improve the accuracy of link prediction.

[0071] In step 11, it is assumed that the preset first data set is the HMDD v2.0 data set, which contains 5430 pairs of known miRNA-disease associations, involving 495 miRNAs and 383 diseases. Among them, 495 miRNAs can be used as 495 miRNA nodes, and 383 diseases can be used as 383 disease nodes. The 5430 pairs of known miRNA-disease associations are used as the association information between nodes. MicroRNA (miRNA) is a small non-coding RNA molecule that plays a vital role in the regulation of gene expression. They promote the degradation of targets or inhibit their translation into proteins by binding to targets. Studies have shown that miRNA is involved in a wide range of physiological and pathological processes, including the generation and development of various diseases. The prediction of miRNA-disease association aims to predict the occurrence of a certain disease by using the expression level of miRNA or other information.

[0072] In step 11, an edge between the nodes may be formed according to the association information between the nodes. An edge refers to an edge connecting only two vertices.

[0073] In step 12, in an exemplary embodiment, the initial feature vectors of the N nodes may be generated according to the node attribute information. For example, the determining of the initial feature vectors of the N nodes may include:

[0074] Determine the disease semantic similarity matrix and miRNA functional similarity matrix;

[0075] Constructing a miRNA-disease adjacency matrix, and determining a disease Gaussian spectrum kernel similarity matrix and a miRNA Gaussian spectrum kernel similarity matrix respectively according to the miRNA-disease adjacency matrix;

[0076] Integrating the disease semantic similarity matrix and the disease Gaussian spectrum kernel similarity matrix to obtain a disease integrated similarity matrix; integrating the miRNA functional similarity matrix and the miRNA Gaussian spectrum kernel similarity matrix to obtain a miRNA integrated similarity matrix;

[0077] The initial feature vectors of the N nodes are determined according to the disease integrated similarity matrix and the miRNA integrated similarity matrix.

[0078] In an exemplary embodiment, the disease semantic similarity matrix SD can be determined as follows:

[0079] Obtaining a medical subject heading (MeSH) database, and constructing a directed acyclic graph for describing the association between diseases according to the MeSH database;

[0080] Determine, according to the directed acyclic graph, a semantic contribution of one disease to another disease in the directed acyclic graph; determine, according to the semantic contribution of the one disease to the other disease, a semantic similarity between the one disease and the other disease;

[0081] A semantic similarity matrix between the diseases is constructed according to the semantic similarities between the diseases.

[0082] Exemplarily, the preset second data set may be a medical subject heading (MeSH) database.

[0083] Directed acyclic graph (DAG) is used to describe the relationship between diseases. For a given disease d(i), its directed acyclic graph can be expressed as DAG(d(i)) = (d(i), D(d(i)), E(d(i))), where D(d(i)) represents the set of disease node d(i) itself and its ancestor nodes, and E(d(i)) contains the set of edges corresponding to disease d(i) and its ancestor nodes. The semantic contribution of disease d(t) to d(i) is as follows:

[0084]

[0085] Therefore, the semantic similarity SD of diseases d(i) and d(j) can be defined as:

[0086]

[0087] In an exemplary embodiment, the miRNA function similarity matrix SM can be determined according to the existing method MISIM, or the calculated miRNA-miRNA function similarity matrix can be directly downloaded from the website (http: / / www.cuilab.cn / fles / images / cuilab / misim.zip).

[0088] In an exemplary embodiment, the Gaussian spectral kernel similarity matrix GKD between diseases can be determined as follows:

[0089] The Gaussian spectral kernel similarity between the corresponding diseases is determined according to the Gaussian similarity between each column vector in the second adjacency matrix; wherein the second adjacency matrix is ​​a matrix constructed according to the association between miRNA and disease, the number of rows is equal to the number of miRNA, and the number of columns is equal to the number of diseases, and the known miRNA-disease association data set is traversed, and for each pair of miRNA-disease associations, the corresponding row and column are found in the matrix, and the value of the position is set to 1, and the values ​​of the remaining positions are set to 0;

[0090] A Gaussian spectrum kernel similarity matrix of the disease is constructed according to the Gaussian spectrum kernel similarity between each disease.

[0091] Exemplarily, the Gaussian spectral kernel similarity GKD(d(i), d(j)) between disease d(i) and disease d(j) is calculated according to the following formula:

[0092] GKD(d(i),d(j))=exp(-γ d ||BV(d(i))-BV(d(j))|| 2 ) (4);

[0093]

[0094] Where BV(d(i)) represents a binary vector, i.e., the i-th column of the second adjacency matrix. BV(d(j)) represents the j-th column of the second adjacency matrix. d Represents the kernel bandwidth parameter. n d Represents the number of all diseases.

[0095] In an exemplary embodiment, the Gaussian kernel similarity matrix GKM of the miRNA can be determined as follows:

[0096] Determine the Gaussian spectral kernel similarity between corresponding miRNAs according to the Gaussian similarity between each row vector in the second adjacency matrix;

[0097] A Gaussian spectrum kernel similarity matrix of the miRNA is constructed according to the Gaussian spectrum kernel similarity between each miRNA.

[0098] Exemplarily, the Gaussian kernel similarity matrix GKM of the miRNA is calculated according to the following formula:

[0099] GKM(m(i),m(j))=exp(-γ m ||BV(m(i))-BV(m(j))|| 2 ) (6);

[0100]

[0101] Where BV(m(i)) represents a binary vector, i.e., the i-th row of the second adjacency matrix. BV(m(j)) represents the j-th row of the second matrix. m is the kernel bandwidth parameter.

[0102] Exemplarily, the above similarity matrices can be integrated according to the following formula:

[0103] For miRNA nodes, the integrated calculation method is as follows:

[0104]

[0105] Among them, SM(m(i),m represents the functional similarity between miRNA m(i) and miRNA m(j), and GKM(m(i),m(j)) represents the Gaussian kernel similarity between miRNAm(i) and miRNAm(j).

[0106] For disease nodes, the integration calculation method is as follows:

[0107]

[0108] Among them, SD(d(i), d(j)) represents the semantic similarity between diseases d(i) and d(j), and GKD(d(i), d(j)) represents the Gaussian spectral kernel similarity between diseases d(i) and d(j).

[0109] In some other exemplary embodiments, the initial feature vectors of the N nodes may be generated by random initialization, but the feature vectors generated in this way do not contain valid node attribute information, so the validity of the subsequently constructed hypergraph will also be affected.

[0110] In step 13, a hyperedge can connect two or more vertices.

[0111] Generating hyperedges between nodes according to the initial feature vectors of the N nodes may include:

[0112] Clustering the initial feature vectors of the N nodes using a K-nearest neighbor algorithm;

[0113] Generate hyperedges between nodes based on the clustering results.

[0114] Exemplarily, clustering the initial feature vectors of the N nodes by using a K-nearest neighbor algorithm may include:

[0115] The Euclidean distance between nodes is calculated using the K nearest neighbor algorithm to achieve node clustering.

[0116] In other exemplary embodiments, the Manhattan distance or Minkowski distance or other distances between nodes may be calculated by using a K-nearest neighbor algorithm to achieve node clustering.

[0117] Among them, the K-Nearest Neighbors (KNN) algorithm is an algorithm that finds the K neighbor samples closest to the target sample in the data set and predicts the label (for classification problems) or value (for regression problems) of the target sample based on the labels of these neighbor samples.

[0118] Exemplarily, generating a hyperedge between nodes according to the clustering result means generating a hyperedge according to the nodes clustered into one category.

[0119] In step 13, in an exemplary embodiment, generating a hyperedge between nodes according to the known edge information between the nodes may include the following steps 131-135:

[0120] Step 131, constructing a common graph according to the known edge information between the nodes;

[0121] Step 132, determining the adjacency matrix of the general graph;

[0122] Step 133, establishing a linear regression model for each node according to the adjacency matrix;

[0123] Step 134, solving the optimization target of the linear regression model of each node;

[0124] Step 135, generating a hyperedge according to the solution result.

[0125] In steps 131-132, illustratively, assuming that the common graph G constructed based on the association between known miRNA-diseases has a total of 878 nodes (495 miRNAs are taken as 495 nodes, and 383 diseases are taken as 383 nodes), the association between the nodes in the graph G can be described by constructing an 878*878 adjacency matrix B. When there is a known association between node i and node j, B(i,j)=1, otherwise the value is 0. The adjacency matrix B is a symmetric matrix that represents the structural information of the known miRNA-disease association network. The i-th row B(i) in the adjacency matrix B represents the structural information of node i in the association network, that is, the association with other nodes.

[0126] In step 133, the linear regression model may be:

[0127] B(i)=X(i)Z(i)+b(i) (10);

[0128] Among them, B represents the adjacency matrix, B(i) is the i-th row of the adjacency matrix, representing the structural information of node i, X(i) represents the structural information of other nodes except node i, Z(i) is the regression coefficient, representing the correlation coefficient between node i and other nodes, and b(i) is the random error term, 1≤i≤N.

[0129] By representing the structural information of node i as a linear combination of the structural information of other nodes, the high-order relationship between node i and other nodes can be reflected. In the above formula (1), the difference between B(i) and X(i)Z(i) should be as small as possible, and in order to avoid overfitting, the size of Z(i) should be as small as possible.

[0130] In step 134, the optimization target can be determined according to the following formula:

[0131] min Z(i) ||B(i)-X(i)Z(i)||+λ||Z(i)|| (11);

[0132] Among them, ||·|| represents the L1 norm, min represents the minimum value, and λ is the regularization parameter, 0<λ<1. λ determines the number of zero elements in Z(i). The larger λ is, the more zero elements there are in Z(i), and vice versa.

[0133] In step 135, generating a hyperedge according to the solution result includes:

[0134] Generate a hyperedge between the node corresponding to the non-zero element in Z(i) and node i.

[0135] In step 14, constructing a hypergraph according to the N nodes and the generated hyperedges may include:

[0136] Construct the N nodes, the hyperedges generated based on the initial features of the nodes in step 13, and the hyperedges generated based on the known edge connection information between the nodes into a hypergraph as the initial hypergraph;

[0137] An example of the constructed hypergraph is as Figure 2 shown, where E1, E2, E3, and E4 represent hyperedges, and n1, n2, n3, n4, n5, n6, and n7 represent nodes.

[0138] In an exemplary embodiment, the hypergraph representation learning for the hypergraph may include:

[0139] Perform hypergraph representation learning and structure adjustment on the initial hypergraph through a multi-layer hypergraph convolutional neural network;

[0140] Each layer of the multi-layer hypergraph convolutional neural network is composed of one layer of a hypergraph convolutional neural network.

[0141] One layer of a hypergraph convolutional neural network (HyperGraph Convolutional Network, HGCN) is a neural network composed of one layer of a hypergraph convolutional layer. The hypergraph convolutional layer is used to perform convolutional processing on the input node features through a convolutional matrix, so as to achieve the effect of feature extraction for the nodes. Feature extraction is performed on the hypergraph through the hypergraph convolutional neural network to obtain the node feature representations (i.e., the feature vectors of the nodes) of each node, and then downstream tasks are performed based on the node feature representations, such as node classification based on the node feature representations, etc.

[0142] In an exemplary embodiment, the performing hypergraph representation learning and structure adjustment on the hypergraph through a multi-layer hypergraph convolutional neural network may include:

[0143] Input the initial hypergraph into the first layer of the multi-layer hypergraph convolutional neural network to obtain the feature vectors of the nodes in the first layer;

[0144] The input of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the (I + 1)-th layer hypergraph generated based on the feature vectors of the nodes in the I-th layer; the output of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the feature vectors of the nodes in the (I + 1)-th layer (as Figure 3 shown), where 1 ≤ I < M, and M is the total number of layers of the multi-layer hypergraph convolutional neural network.

[0145] In this embodiment, a multi-layer hypergraph convolutional neural network is used to perform learning and structure adjustment on the hypergraph, which can enable the hypergraph structure to be continuously adjusted as the model is trained. Compared with existing methods, it can fully explore the real hypergraph structure and better mine the potential high-order relationships between data.

[0146] In an exemplary embodiment, generating the hypergraph of the I+1th layer according to the feature vectors of the nodes of the Ith layer may include:

[0147] The hypergraph is reconstructed based on the feature vectors of the nodes in the I layer based on the K nearest neighbor algorithm to generate the I+1 layer hypergraph.

[0148] This embodiment not only generates hyperedges based on the initial feature vectors of the nodes, but also generates hyperedges based on the association information between the nodes, and constructs a hypergraph based on the nodes and the generated hyperedges, so that the constructed hypergraph not only reflects the high-order relationship between the node attributes, but also reflects the high-order relationship of the known network structure, and performs hypergraph learning and dynamic structure adjustment based on a multi-layer hypergraph convolutional neural network, and performs link prediction based on the final node representation vector, which further helps to improve the accuracy of link prediction.

[0149] In step 15, performing link prediction between nodes according to the hypergraph representation learning result may include:

[0150] Calculating the cosine similarity between the feature vectors of the node pairs to be predicted based on the hypergraph representation learning result, and judging whether there is a correlation between the node pairs to be predicted based on a preset cosine similarity threshold;

[0151] or,

[0152] Calculate the inner product between the feature vectors of the node pair to be predicted according to the hypergraph representation learning result, and determine whether there is a correlation between the node pair to be predicted according to a preset inner product threshold;

[0153] or,

[0154] The feature vectors of the node pairs to be predicted are concatenated according to the hypergraph representation learning result, and the concatenated feature vectors are input into a binary classifier to determine whether there is a correlation between the node pairs to be predicted.

[0155] In an exemplary embodiment, the hypergraph representation learning result is a feature vector of a node of the Mth layer;

[0156] Exemplarily, a specific method for performing link prediction between nodes according to the feature vectors of the nodes in the Mth layer is as follows:

[0157] Calculate the cosine similarity between the feature vectors of the node pairs to be predicted based on the feature vectors of the nodes in the Mth layer, and determine whether there is a correlation between the node pairs to be predicted based on a preset cosine similarity threshold;

[0158] or,

[0159] Calculate the inner product between the feature vectors of the node pairs to be predicted based on the feature vectors of the nodes in the Mth layer, and determine whether there is a correlation between the node pairs to be predicted based on a preset inner product threshold;

[0160] or,

[0161] The feature vectors of the node pairs to be predicted are concatenated according to the feature vectors of the nodes in the Mth layer, and the concatenated feature vectors are input into a binary classifier to determine whether there is a correlation between the node pairs to be predicted.

[0162] Figure 4 Schematic diagram of a miRNA-disease association prediction method according to an embodiment of the present application, such as Figure 4 As shown, the miRNA-disease association prediction method of the embodiment of the present application includes the following steps 41 to 43.

[0163] Step 41, obtaining a node pair to be tested;

[0164] Step 42, determining a representation vector of each node according to the miRNA-disease hypergraph;

[0165] Step 43, performing vector splicing on the node pairs to be tested according to the representation vector of each node, and performing association prediction according to the spliced ​​vectors;

[0166] Wherein, the node pairs to be tested include miRNA-disease pairs;

[0167] The miRNA-disease supergraph is constructed as follows:

[0168] The dataset containing miRNA and associated diseases is used as the preset first dataset;

[0169] A matrix obtained by integrating a disease-disease semantic similarity matrix, a miRNA-miRNA functional similarity matrix, a Gaussian spectral kernel similarity matrix between diseases, and a Gaussian spectral kernel similarity matrix between miRNAs is used as an initial feature vector of N nodes; wherein the N nodes include miRNA nodes and disease nodes;

[0170] Using a first adjacency matrix as known network structure information; wherein the first adjacency matrix is ​​an adjacency matrix of a common graph formed according to the association between miRNA and disease;

[0171] The miRNA-disease supergraph was constructed according to the aforementioned supergraph construction method.

[0172] The miRNA-disease association prediction method of the embodiment of the present application generates hyperedges based on multiple information, including node features and known network structure information, and then constructs a hypergraph based on the nodes and hyperedges, and performs miRNA-disease association prediction through the constructed hypergraph. Since the topological structure information of the known miRNA-disease association network is fully utilized, the high-order relationship of the network structure information can be better reflected, so that the constructed hypergraph contains richer information, which helps to improve the accuracy of miRNA-disease association prediction.

[0173] In step 42, illustratively, the dataset containing miRNA-associated disease pairs may be the HMDD v2.0 dataset, which contains 5430 pairs of known miRNA-disease associations, involving 495 miRNAs and 383 diseases.

[0174] In an exemplary embodiment, constructing a miRNA-disease hypergraph according to the hypergraph construction method described in any of the above embodiments includes generating hyperedges between miRNA-miRNA nodes, disease-disease nodes, and miRNA-disease nodes according to node attribute information, and generating hyperedges between miRNA-miRNA nodes, disease-disease nodes, and miRNA-disease nodes according to known network structure information.

[0175] Among them, the hyperedges between miRNA-miRNA nodes, between disease-disease nodes, and between miRNA-disease nodes are generated according to the node attribute information, including:

[0176] According to the node attribute information, the Euclidean distances between nodes were calculated according to the K nearest neighbor algorithm, and the nodes were clustered according to the Euclidean distances between nodes to generate hyperedges between miRNA-miRNA nodes, between disease-disease nodes, and between miRNA-disease nodes.

[0177] Node attribute information can be represented as the initial feature vector of the node.

[0178] Hyperedges generated using node attributes focus more on reflecting the high-order relationships of node attributes in the network.

[0179] Generate hyperedges between miRNA-miRNA nodes, disease-disease nodes, and miRNA-disease nodes based on known network structure information, including:

[0180] Establish the first adjacency matrix;

[0181] For each node, perform the following operations:

[0182] The association information between the node and other nodes in the adjacency matrix is ​​used as the network structure information of the node; the network structure information of the node is represented by a linear regression model; wherein the linear regression model includes a regression coefficient; the optimization target of the linear regression model is solved to obtain the regression coefficient; and a hyperedge is generated between the node and the node according to the node corresponding to the element that is not 0 in the regression coefficient.

[0183] The first adjacency matrix is ​​an adjacency matrix of a common graph formed according to the associations among miRNA nodes, disease nodes and known miRNA-diseases.

[0184] The above-mentioned hyperedge construction method based on network structure information makes full use of the topological structure information of the known miRNA-disease association network and can better reflect the high-order relationship of the network structure information.

[0185] After constructing a miRNA-disease hypergraph based on miRNA nodes, disease nodes, hyperedges between miRNA-miRNA nodes, disease-disease nodes, and miRNA-disease nodes according to node attribute information, and hyperedges between miRNA-miRNA nodes, disease-disease nodes, and miRNA-disease nodes according to known network structure information, the miRNA-disease hypergraph can also be trained based on a multi-layer hypergraph convolutional neural network. The representation vectors of miRNA nodes and disease nodes (the feature matrix of the above-mentioned nodes includes the representation vectors of the nodes) are updated using a hypergraph convolutional neural network at each layer, and the next layer of hypergraph convolutional neural network is clustered according to the updated representation vectors of miRNA nodes and disease nodes to reconstruct the miRNA-disease hypergraph, thereby learning and adjusting the structure of the miRNA-disease hypergraph.

[0186] In step 42, the representation vector of each node can be determined according to the miRNA-disease hypergraph; the miRNA-disease hypergraph here can be either a hypergraph that has not been learned and adjusted, or a hypergraph that has been learned and adjusted.

[0187] In step 43, association prediction is performed based on the concatenated vectors, which may include:

[0188] The spliced ​​vector is input into a bipartitioner, and the bipartitioner is used to predict whether there is a correlation between miRNA and disease in the node pair to be tested.

[0189] The present application also provides a link prediction device (such as Figure 5 ), including a memory and a processor,

[0190] The memory is used to store a program for link prediction;

[0191] The processor is used to read the program for executing the link prediction and execute the link prediction method described in any one of the above embodiments.

[0192] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to enable the computer to execute the link prediction method described in any of the above embodiments.

[0193] The present application describes multiple embodiments, but the description is exemplary rather than restrictive, and it is obvious to those skilled in the art that there may be more embodiments and implementations within the scope of the embodiments described in the present application. Although many possible feature combinations are shown in the drawings and discussed in the specific embodiments, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with any other feature or element in any other embodiment, or may replace any other feature or element in any other embodiment.

[0194] The present application includes and contemplates combinations of features and elements known to those of ordinary skill in the art. The embodiments, features and elements disclosed in the present application may also be combined with any conventional features or elements to form a unique invention scheme. Any features or elements of any embodiment may also be combined with features or elements from other invention schemes to form another unique invention scheme. Therefore, it should be understood that any feature shown and / or discussed in the present application may be implemented individually or in any appropriate combination. Therefore, except for the limitations made according to the attached claims and their equivalents, the embodiments are not subject to other restrictions. In addition, various modifications and changes may be made within the scope of protection of the attached claims.

[0195] In addition, when describing representative embodiments, the specification may have presented the method and / or process as a specific sequence of steps. However, to the extent that the method or process does not rely on the specific order of the steps described herein, the method or process should not be limited to the steps of the specific order described. As will be understood by those of ordinary skill in the art, other sequences of steps are also possible. Therefore, the specific sequence of the steps set forth in the specification should not be interpreted as a limitation to the claims. In addition, the claims for the method and / or process should not be limited to the steps of performing them in the order written, and those skilled in the art can easily understand that these sequences can be changed and still remain within the spirit and scope of the embodiments of the present application.

[0196] The functional modules / units in all or some steps, systems, and devices disclosed in the above method may be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those of ordinary skill in the art, the term "computer storage medium" includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0197] In addition, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first", "second", etc. may explicitly or implicitly include at least one of the features.

[0198] In the description of the present application, “plurality” means at least two, for example, two, three, etc., unless otherwise clearly and specifically defined.

[0199] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0200] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A link prediction method, characterized in that, it includes: obtaining a preset first data set, determining N nodes and the known edge connection information between the nodes according to the preset first data set, where N is a positive integer greater than or equal to 3; determining the initial feature vectors of the N nodes; generating hyperedges between the nodes respectively according to the initial feature vectors of the N nodes and the known edge connection information between the nodes; constructing a hypergraph based on the N nodes and the generated hyperedges and performing hypergraph representation learning on the hypergraph; performing link prediction between the nodes according to the hypergraph representation learning result.

2. The link prediction method according to claim 1, characterized in that: generating hyperedges between the nodes according to the known edge connection information between the nodes includes: constructing an ordinary graph according to the known edge connection information between the nodes; determining the adjacency matrix of the ordinary graph; establishing a linear regression model for each node according to the adjacency matrix; solving the optimization objective of the linear regression model of each node; generating hyperedges according to the solution result.

3. The link prediction method according to claim 2, characterized in that: the linear regression model is: B(i) = X(i)Z(i) + b(i); where B represents the adjacency matrix, B(i) is the i-th row of the adjacency matrix, representing the structural information of node i, X(i) represents the structural information of other nodes except node i, Z(i) is the regression coefficient, representing the correlation coefficient between node i and other nodes, and b(i) is the random error term, 1 ≤ i ≤ N.

4. The link prediction method according to claim 3, characterized in that: the optimization objective is determined according to the following formula: Minutes Z(i) ||B(i)-X(i)Z(i)||+λ||Z(i)||; where ||·|| represents the L1 norm, min represents taking the minimum value, and λ is the regularization parameter, 0 < λ < 1.

5. The link prediction method according to claim 4, characterized in that: generating hyperedges according to the solution result includes: generating a hyperedge between the node corresponding to the non-zero element in Z(i) and node i.

6. The link prediction method according to claim 1, characterized in that: performing hypergraph representation learning on the hypergraph includes: taking the hypergraph constructed by the N nodes and the generated hyperedges as the initial hypergraph; performing hypergraph representation learning and structure adjustment on the initial hypergraph through a multi-layer hypergraph convolutional neural network; where each layer of the multi-layer hypergraph convolutional neural network is composed of one layer of hypergraph convolutional neural network.

7. The link prediction method according to claim 6, characterized in that: performing hypergraph representation learning and structure adjustment on the hypergraph through a multi-layer hypergraph convolutional neural network includes: inputting the initial hypergraph into the first layer of the multi-layer hypergraph convolutional neural network to obtain the feature vectors of the nodes in the first layer; the input of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the (I + 1)-th layer hypergraph generated according to the feature vectors of the nodes in the I-th layer; the output of the (I + 1)-th layer of the multi-layer hypergraph convolutional neural network is the feature vectors of the nodes in the (I + 1)-th layer, where 1 ≤ I < M, and M is the total number of layers of the multi-layer hypergraph convolutional neural network.

8. The link prediction method according to claim 7, Features: The step of generating the hypergraph of the I+1th layer according to the feature vectors of the nodes of the Ith layer comprises: The I+1th layer hypergraph is generated based on the K nearest neighbor algorithm according to the feature vectors of the nodes in the Ith layer.

9. The link prediction method according to claim 7, Features: The hypergraph representation learning result is the feature vector of the node of the Mth layer.

10. The link prediction method according to any one of claims 1, 6 to 7, Features: The link prediction between nodes according to the hypergraph representation learning result includes: Calculating the cosine similarity between the feature vectors of the node pairs to be predicted based on the hypergraph representation learning result, and judging whether there is a correlation between the node pairs to be predicted based on a preset cosine similarity threshold; or, Calculate the inner product between the feature vectors of the node pair to be predicted according to the hypergraph representation learning result, and determine whether there is a correlation between the node pair to be predicted according to a preset inner product threshold; or, The feature vectors of the node pairs to be predicted are concatenated according to the hypergraph representation learning result, and the concatenated feature vectors are input into a binary classifier to determine whether there is a correlation between the node pairs to be predicted.

11. The link prediction method according to claim 1, Features: The N nodes include miRNA nodes and disease nodes; The determining of the initial feature vectors of the N nodes comprises: Determine the disease semantic similarity matrix and miRNA functional similarity matrix; Constructing a miRNA-disease adjacency matrix, and determining a disease Gaussian spectrum kernel similarity matrix and a miRNA Gaussian spectrum kernel similarity matrix respectively according to the miRNA-disease adjacency matrix; Integrating the disease semantic similarity matrix and the disease Gaussian spectrum kernel similarity matrix to obtain a disease integrated similarity matrix; integrating the miRNA functional similarity matrix and the miRNA Gaussian spectrum kernel similarity matrix to obtain a miRNA integrated similarity matrix; The initial feature vectors of the N nodes are determined according to the disease integrated similarity matrix and the miRNA integrated similarity matrix.

12. A link prediction device, comprising a memory and a processor, It is characterized in that The memory is used to store a program for link prediction; The processor is used to read the program for executing the link prediction and execute the method according to any one of claims 1 to 11.

13. A computer-readable storage medium storing computer-executable instructions, in, The computer executable instructions are used to cause the computer to execute the method according to any one of claims 1 to 11.

Images