Method and device for network alignment

Abstract

The present invention relates to a method and a device for network alignment using only a graph structure in a situation where node attribute information is not present. The method for network alignment, related to one embodiment of the present invention, may comprise the steps of: acquiring structural relationship information for each of nodes constituting a source network and a target network, wherein the source network and the target network comprise an anchor node; for each of the source network and the target network, measuring the structural similarity between the anchor node and the respective nodes constituting the network; generating an attribute value for each of the nodes by using the measured structural similarity, assigning the generated attribute value to each of the nodes, and vectorizing each of the nodes to which the respective attribute values have been assigned; and aligning the source network and the target network according to the similarity between the respective vectorized nodes of the source network and the target network.

Description

Network alignment method and device

[0001] The present invention relates to a network alignment method and apparatus, and more specifically, to a method and apparatus for aligning a network using only a graph structure in a situation where node attribute information is unavailable.

[0002] Network data (or graph data) is highly useful for describing various objects and the relationships between them, such as social networks, relational networks, molecular structures, and recommendation systems. Network data (hereinafter referred to as 'network') consists of nodes corresponding to each object and edges connecting nodes according to the relationships between the nodes, enabling the analysis of associations between multiple objects.

[0003] For example, in a social network, nodes can represent users and edges can represent relationships between users (e.g., friends), while in a relational network, nodes can represent individual papers and edges can represent citation relationships. Furthermore, in a recommendation system, nodes can be users or products, and edges can represent recommendation relationships.

[0004] Meanwhile, a multiple network refers to a network in which distinct networks include nodes for at least one identical object. Multiple networks are utilized in a wide range of application fields, from computer vision, bioinformatics, and web mining to chemistry and social network analysis.

[0005] In multi-networks, it is crucial to identify nodes related to the same object across multiple nodes within each network. The process of identifying corresponding nodes across different networks is called Network Alignment (NA) (also known as graph matching). In other words, network alignment refers to the task of detecting corresponding nodes between two different networks based on the structure and node attributes of each network. Network alignment can be utilized as an initial step for downstream machine learning tasks across multiple networks. For example, identifying different accounts of the same user on various social networks (e.g., Facebook, Twitter, etc.) through network alignment facilitates friend recommendations, user behavior prediction, and personalized advertising. Furthermore, in bioinformatics, aligning specific protein-protein interaction (PPI) networks allows for the effective prioritization of candidate genes.

[0006] However, when node attribute information was unavailable, there were limitations to network alignment, which is the task of detecting corresponding nodes between two different networks.

[0007] Therefore, there is a need for research on techniques to perform network alignment even when node attribute information is unavailable.

[0008]

[0009] The objective of the present invention is to provide a network alignment method and apparatus that enable network alignment to be performed even when there is no attribute information of the nodes.

[0010] The objective of the present invention is to provide a network alignment method and apparatus that, when attribute information of a node is unavailable, measure the structural similarity between each node constituting a network and an anchor node, and generate attribute values ​​for each node based on the measured structural similarity to perform network alignment.

[0011] According to one embodiment of the present invention to achieve the above objective, a network alignment method is disclosed, comprising: a step of obtaining structural relationship information for each of the nodes constituting a source network and a target network, wherein the source network and the target network include an anchor node; a step of measuring structural similarity between each node constituting each network and the anchor node for each of the source network and the target network; a step of generating an attribute value for each node using the measured structural similarity, assigning the generated attribute value to each node, and vectorizing each node to which the attribute value is assigned; and a step of aligning the source network and the target network according to the similarity between the vectorized nodes of each of the source network and the target network.

[0012] According to an embodiment of the present invention, to achieve the above objective, a network alignment device is disclosed, comprising a device having one or more processors and a memory for storing one or more programs executed by said one or more processors, wherein the processor acquires structural relationship information for each of the nodes constituting a source network and a target network, measures the structural similarity between each node constituting each network and an anchor node for each of said source network and the target network, generates an attribute value for said node using said structural similarity, assigns the generated attribute value to said node, vectorizes said node to which the attribute value is assigned, and aligns said source network and target network according to the similarity between the vectorized nodes of said source network and target network.

[0013]

[0014] A network alignment method and device according to one embodiment of the present invention can perform network alignment even when there is no attribute information of the nodes.

[0015] According to one embodiment of the present invention, when there is no attribute information of a node, an attribute value of each node is generated based on the measured structural similarity between each node constituting the network and an anchor node, and the attribute value is optimally utilized to improve network alignment performance.

[0016]

[0017] FIG. 1 is a diagram illustrating the concept of a network alignment device aligning a network in relation to an embodiment of the present invention.

[0018] FIG. 2 is a block diagram showing a configuration classified according to operation in a network alignment device related to an embodiment of the present invention.

[0019] FIG. 3 is a diagram illustrating a method for configuring a similarity matrix in a network alignment device related to an embodiment of the present invention.

[0020] FIG. 4 is a diagram illustrating a method for assigning attributes to nodes in a network alignment device related to an embodiment of the present invention.

[0021] FIG. 5 is a diagram illustrating the process of outputting an embedding by reflecting a correction factor in a network alignment device related to an embodiment of the present invention.

[0022] Figure 6 is a detailed block diagram of the similarity calculation unit illustrated in Figure 2.

[0023] FIG. 7 is a diagram showing a pair of nodes finally aligned in a network alignment device related to an embodiment of the present invention.

[0024] FIG. 8 is a diagram illustrating a computing environment including a computing device related to an embodiment of the present invention.

[0025]

[0026] In order to fully understand the present invention, the operational advantages of the present invention, and the objectives achieved by the implementation of the present invention, reference must be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

[0027] The present invention will be described in detail below by explaining preferred embodiments with reference to the attached drawings. However, the present invention may be implemented in various different forms and is not limited to the embodiments described. Furthermore, to clearly explain the present invention, parts unrelated to the description are omitted, and the same reference numerals in the drawings indicate the same components.

[0028] The embodiments and the terms used therein are not intended to limit the technology described in this document to specific embodiments and should be understood to include various modifications, equivalents, and / or substitutions of said embodiments.

[0029] In describing various embodiments below, if it is determined that a detailed description of related known functions or configurations could unnecessarily obscure the essence of the invention, such detailed description will be omitted.

[0030] In relation to the description of the drawings, similar reference numerals may be used for similar components.

[0031] A singular expression may include a plural expression unless the context clearly indicates otherwise.

[0032] In this document, expressions such as "A or B" or "at least one of A and / or B" may include all possible combinations of the items listed together.

[0033] Where it is stated that a certain (e.g., first) component is "(functionally or telecommunicationally) connected" or "connected" to another (e.g., second) component, the certain component may be directly connected to the other component or connected through another component (e.g., third component).

[0034] As used in this specification, singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “composed” or “comprising” should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as meaning that some of the components or steps may not be included, or that additional components or steps may be included.

[0035] In this specification, network data (or graph data) (hereinafter referred to as "network") is data that is highly useful for describing various objects and the relationships between objects, such as social networks, relational networks, molecular structures, and recommendation systems. A network consists of nodes corresponding to each object and edges connecting nodes according to the relationships between the nodes, enabling the analysis of associations between multiple objects.

[0036] FIG. 1 is a diagram illustrating the concept of a network alignment device aligning a network in relation to an embodiment of the present invention.

[0037] Figure 1 shows network alignment performed using an artificial neural network, and (b) shows network alignment to be performed in this embodiment.

[0038] As illustrated in FIG. 1, network alignment using an artificial neural network in the past was performed by the artificial neural network performing neural network operations on a source network (Gs) and a target network (Gt) to detect all pairs of aligned nodes that can be aligned (or matched) at once among a number of source nodes included in the source node set (Vs) of the source network (Gs) and a number of target nodes included in the target node set (Vt) of the target network (Gt). At this time, the artificial neural network aligns the network by detecting pairs of aligned nodes in the node set (Vs, Vt) based on the attribute set (Xs, Xt) for the nodes of the node set (Vs, Vt) and the edge set (Es, Et) connecting the nodes of each node set (Vs, Vt).

[0039] However, there were limitations in performing network alignment in environments without node attributes.

[0040] In the following examples, we will describe a method for aligning networks in an environment where different networks (e.g., Gs, Gt) contain anchor nodes, but there are no attribute values ​​(or attribute information) for each node constituting each network. An anchor node refers to a node used as an important reference point in graph or network analysis. It is primarily used to find pairs of nodes with the same meaning in alignment or matching tasks between multiple networks. These nodes represent information common to different networks.

[0041] In this specification, an anchor node refers to a node configured as the same user or entity in two or more networks.

[0042] FIG. 2 is a block diagram showing a configuration classified according to operation in a network alignment device related to an embodiment of the present invention.

[0043] As described, the network alignment device (100) may include a network acquisition unit (10), a structural similarity measurement unit (20), an attribute value assignment unit (30), an embedding output unit (40), a similarity calculation unit (50), and an alignment unit (60).

[0044] The network acquisition unit (10) acquires two networks (Gs, Gt) that need to be aligned. One of the two networks (Gs, Gt) is called the source network (Gs), and the other is called the target network (Gt).

[0045] And each of the two networks (Gs, Gt) includes multiple nodes and may include structural relationship information representing the connection structure between multiple nodes in each network (Gs, Gt).

[0046] The structural similarity measurement unit (20) can measure the structural similarity between each node constituting each network and the anchor node for each of the two networks (Gs, Gt). The structural similarity measurement can be performed based on the graph connection relationships of each of the two networks (Gs, Gt).

[0047] Structural similarity measurement can refer to a method of measuring how close two nodes are within a graph by considering the topology within a given graph structure. Structural similarity can be measured by multiple methods.

[0048] Structural similarity measurement methods can determine whether a given pair of nodes is close by considering the structure of each node and its adjacent nodes (immediate vicinity) (e.g., Jaccard similarity, AdarSim similarity, etc.), or by considering a broader range of structures (e.g., Simrank and Random Walk with Restart, etc.).

[0049] The structural similarity measuring unit (20) can construct a similarity matrix based on the measured structural similarity. Below, in the embodiment, a method for constructing a similarity matrix based on structural similarity measured in a plurality of ways will be described.

[0050] FIG. 3 is a diagram illustrating a method for configuring a similarity matrix in a network alignment device related to an embodiment of the present invention.

[0051] First, the network acquisition unit (10) acquires two networks (Gs, Gt) that need to be aligned (S310). Each of the two networks (Gs, Gt) includes multiple nodes and may include structural relationship information indicating the connection structure between multiple nodes in each network (Gs, Gt).

[0052] The structural similarity measurement unit (20) measures the structural similarity between each node constituting each network and the anchor node for each of the two networks (Gs, Gt) using a plurality of methods, and can construct a similarity matrix based on the measured structural similarity (S320).

[0053] In the illustrated embodiment, I will describe an example of constructing a similarity matrix based on structural similarity measured by three methods (1st Structural Relationship, 2nd Structural Relationship, 3rd Structural Relationship).

[0054] In the above similarity matrix, the number of either row or column can be equal to the number of anchor nodes, and the number of the other row or column can be equal to the number of nodes constituting each network.

[0055] In the illustrated embodiment, the columns were made equal to the number of anchor nodes (X, Y, Z, x, y, z), and the rows were made equal to the number of nodes constituting each network to form a similarity matrix.

[0056] And the matrix value is the measured structural similarity between the anchor node corresponding to each row and each column and the node constituting each network. For example, the structural similarity between node C and anchor node X in the source network (Gs) is R1.

[0057] In the illustrated embodiment, three similarity matrices (source network: RS1, RS2, RS3 / target network: RT1, RT2, RT3) based on structural similarity measured by three methods (1st Structural Relationship, 2nd Structural Relationship, 3rd Structural Relationship), respectively, can be expanded so that the rows match (so that each node matches) to form a single similarity matrix (similarity matrix in the source network: Ms / similarity matrix in the target network: Mt).

[0058] Referring again to FIG. 2, the attribute value assignment unit (30) can generate an attribute value for each node using the measured structural similarity and assign the generated attribute value to each node.

[0059] FIG. 4 is a diagram illustrating a method for assigning attributes to nodes in a network alignment device related to an embodiment of the present invention.

[0060] The attribute value assignment unit (30) can receive a similarity matrix (Ms) in the source network and a similarity matrix (Mt) in the target network, respectively. Then, it can extract a row corresponding to each node (e.g., a row corresponding to node A) from each of the received similarity matrices (Ms, Mt). For example, if a row corresponding to node A is extracted, a 1x9 vector value can be extracted.

[0061] The attribute value assignment unit (30) uses the matrix value (1x9 vector value) of the extracted row as input to the learned selection neural network (31, 32) to select a specific matrix value by the selection neural network (31, 32), and can treat matrix values ​​not selected by the selection neural network as 0.

[0062] For example, rows (A1, A2, ⪋, A9) corresponding to node A in the source network can be extracted, and the first two of the nine values ​​are selected by the selection neural network (31), while the remaining seven are treated as 0 to form a vector (a vector of 1x9) (A1, A2, 0, 0, ⪋, 0).

[0063] The selection neural network (31) for selecting matrix values ​​of the similarity matrix of the source network and the selection neural network (32) for selecting matrix values ​​of the similarity matrix of the target network may be independent.

[0064] The above selection neural network (31, 32) may use a fully connected layer (Ws, Wt) and a sigmoid layer (σ).

[0065] The attribute value assignment unit (30) can assign a vector value corresponding to each node output through the selection neural network (31, 32) as an attribute value for each node.

[0066] Referring again to FIG. 2, the embedding output unit (40) can output an embedding vector by using as input the connection information (graph information) of each node, in which a vector value corresponding to each node output through the selection neural network (31, 32) is assigned as an attribute value.

[0067] The embedding output unit (40) can project multiple nodes into the embedding space using a network embedding (NE) technique that models high-order connection information of the network (Gs, Gt) into vectors in a low-dimensional virtual embedding space.

[0068] The embedding output unit (40) can be implemented as an artificial neural network. For example, the embedding output unit (40) can be implemented as a Graph Convolutional Network (GCN). A Graph Convolutional Network (GCN) is a neural network designed to process graph-structured data. While traditional neural networks or Convolutional Neural Networks (CNNs) mainly handle Euclidean-structured data such as images, text, and time-series data, a GCN can process non-Euclidean-structured data (graphs) represented by nodes and edges.

[0069] FIG. 5 is a diagram illustrating the process of outputting an embedding by reflecting a correction factor in a network alignment device related to an embodiment of the present invention.

[0070] As described, the embedding output unit (40) can cause similar nodes to be placed close together in the embedding space, while different nodes are placed far apart, by means of a correction factor. A reconstruction loss value may be used as the correction factor. The reconstruction loss value is a correction factor that creates the embedding vector similar to the actual structure of the source network and the target network, and can be expressed in the following Equation 1.

[0071]

[0072] L represents the reconstruction loss, and : Diagonal matrix. is the degree matrix, and * represents the source graph and target graph. The value of the (i,i) element represents the degree information of the i-th node.

[0073] : Adjacency matrix with self-connection. is a matrix obtained by adding the identity matrix I to the adjacency matrix A, which represents the connection information of the existing graph.

[0074] thus, is a matrix representing the probability that two nodes (source, target) are connected in each graph.

[0075] : Hidden representation at GCN layer. This is the output matrix with node embeddings using GCN.

[0076] : It is the Frobenius norm.

[0077] The similarity calculation unit (50) calculates the similarity between nodes output as embedding vectors of the source network and the target network, respectively, and the alignment unit (60) can align the source network and the target network according to the calculated similarity.

[0078] Figure 6 is a detailed block diagram of the similarity calculation unit illustrated in Figure 2.

[0079] As described, the similarity calculation unit (50) may include an embedding similarity unit (51), a Tsversky similarity unit (52), and a weighted similarity unit (53).

[0080] The embedding similarity unit (51) can calculate the embedding similarity between the nodes of the source network and the nodes of the target network. That is, the embedding similarity unit (51) calculates the similarity between the nodes vectorized into embedding vectors.

[0081] The Tversky similarity unit (52) can calculate Tversky similarity to search for the next pair of nodes to be aligned with high accuracy by additionally referencing the relationship of the previously aligned pair of nodes, while suppressing the occurrence of false positives due to the difference in scale between the two networks (Gs, Gt), that is, the difference in the number of nodes of the source network and the target network (ns, nt), while simultaneously suppressing false positives.

[0082] Here, Zversky similarity can be calculated as the ratio of the number of previously sorted nodes among neighbor nodes to the total number of normalized neighbor nodes for each combined node when node pairs are formed from combinations of unsorted nodes in two networks (Gs, Gt).

[0083] The weighted similarity unit (53) can calculate weighted similarity by weighting the embedding similarity and the Tsversky similarity.

[0084] And the alignment unit (60) can align two networks (Gs, Gt) based on the calculated weighted similarity.

[0085] In this case, the two networks (Gs, Gt) can be sorted at once, but only the nodes with the highest similarity can be partially sorted, and the sorting can be performed incrementally using the partially sorted nodes as anchor nodes.

[0086] FIG. 7 is a diagram showing a pair of nodes finally aligned in a network alignment device related to an embodiment of the present invention.

[0087] By the method described above, if a network alignment is performed by assigning attribute values ​​to nodes that have not been assigned attribute values, node pairs can be formed in the form of Fig. 7.

[0088] FIG. 8 is a diagram illustrating a computing environment including a computing device related to an embodiment of the present invention.

[0089] In the illustrated embodiment, each component may have different functions and capabilities in addition to those described below, and may include additional components in addition to those not described below. The illustrated computing environment includes a computing device (200), and the computing device (200) may be one or more components included in the network alignment device illustrated in FIG. 2.

[0090] A computing device (200) may include at least one processor (210) and a memory (220) that stores one or more programs executed by the one or more processors (210).

[0091] The processor (210) can enable the computing device (200) to operate according to the exemplary embodiment mentioned above. For example, the processor (210) can execute one or more programs stored in computer-readable memory (220).

[0092] The above one or more programs may include one or more computer-executable instructions, and the computer-executable instructions may be configured to cause the computing device (200) to perform operations according to exemplary embodiments when executed by the processor (210).

[0093] As described above, the network alignment method and device according to one embodiment of the present invention can perform network alignment even when there is no attribute information of the node.

[0094] According to one embodiment of the present invention, when there is no attribute information of a node, an attribute value of each node is generated based on the measured structural similarity between each node constituting the network and an anchor node, and the attribute value is optimally utilized to improve network alignment performance.

[0095] The network alignment device and method described above are not limited to the configurations and methods of the embodiments described above; rather, all or part of each embodiment may be selectively combined to allow for various modifications to be made.

Claims

1. A step of obtaining structural relationship information for each of the nodes constituting the source network and the target network - the source network and the target network include anchor nodes -; A step of measuring the structural similarity between each node constituting each network and the anchor node for each of the source network and the target network; A step of generating attribute values ​​for each node using the structural similarity measured above, assigning the generated attribute values ​​to each node, and vectorizing each node to which the attribute values ​​are assigned; and A network alignment method characterized by including the step of aligning the source network and the target network according to the similarity between the vectorized nodes of each of the source network and the target network.

2. In Paragraph 1, The step of measuring the structural similarity above includes the step of constructing a similarity matrix, and In the above similarity matrix, the number of rows and columns is equal to the number of anchor nodes, and The number of the other of the above rows and columns is equal to the number of nodes constituting each of the above networks, and A network alignment method characterized in that the matrix value is the structural similarity measured between the anchor node corresponding to each row and each column and the node constituting each network.

3. In Paragraph 2, The step of measuring the structural similarity above A network alignment method characterized by including the step of expanding multiple similarity matrices based on structural similarity measured by multiple methods, such that rows or columns match, to form a single similarity matrix.

4. In paragraph 3, the step of vectorizing each of the above nodes A step of extracting a row or column corresponding to each node constituting each network for each of the source network and the target network; and A network alignment method characterized by including the step of using the matrix values ​​of the extracted rows or columns as input to a learned selection neural network to select specific matrix values ​​by the selection neural network, and treating matrix values ​​not selected by the selection neural network as 0.

5. In paragraph 4, the step of vectorizing each of the above nodes A network alignment method characterized by including the step of outputting a vector of each node based on the selected specific matrix value and the matrix value processed as zero.

6. In paragraph 5, the above-mentioned selection neural network is A network alignment method characterized by being used independently for each of the source network and the target network.

7. In paragraph 5, the step of outputting the vector of each node is The method further includes the step of outputting an embedding vector by using the connection information of each node, to which the vector of each node output above is assigned as an attribute value, as the input to a graph neural network, wherein The above graph neural network outputs the embedding vector of each node through the reconstruction loss value, and A network alignment method characterized in that the above reconstruction loss value is a correction factor that creates the embedding vector similar to the actual structure of the source network and the target network.

8. A device having one or more processors and a memory for storing one or more programs executed by said one or more processors, wherein The above processor Obtain structural relationship information for each of the nodes constituting the source network and the target network, and For each of the above source network and the above target network, the structural similarity between each node constituting each network and the anchor node is measured, and Using the structural similarity measured above, attribute values ​​for each of the above nodes are generated, the generated attribute values ​​are assigned to each of the above nodes, and each of the above nodes to which the attribute values ​​are assigned is vectorized. A network alignment device characterized by aligning the source network and the target network according to the similarity between the vectorized nodes of each of the source network and the target network.

9. In paragraph 8, the above processor The step of measuring the structural similarity above involves constructing a similarity matrix, and In the above similarity matrix, the number of rows and columns is equal to the number of anchor nodes, and The number of the other of the above rows and columns is equal to the number of nodes constituting each of the above networks, and A network alignment device characterized in that the matrix value is the structural similarity measured between the anchor node corresponding to each row and each column and the node constituting each network.

10. In paragraph 9, the processor A network alignment device characterized by expanding multiple similarity matrices based on structural similarity measured by multiple methods, such that rows or columns match, to form a single similarity matrix.

11. In Clause 10, the above processor For each of the above source network and the above target network, a row or column corresponding to each node constituting each network is extracted, and A network alignment device characterized by using the matrix values ​​of the extracted rows or columns as input to a learned selection neural network to select specific matrix values ​​by the selection neural network, and treating matrix values ​​not selected by the selection neural network as 0.

12. In paragraph 11, the above processor A network alignment device characterized by outputting a vector of each node based on the selected specific matrix value and the matrix value processed as zero.

13. In Clause 12, the above-mentioned selection neural network is A network alignment device characterized by being used independently for each of the source network and the target network.

14. In Clause 12, the above processor The vector of each node output above is used as an input to a graph neural network to output an embedding vector, using the connection information of each node assigned as an attribute value. The above graph neural network outputs the embedding vector of each node through the reconstruction loss value, and A network alignment device characterized in that the above reconstruction loss value is a correction factor that creates the embedding vector similar to the actual structure of the source network and the target network.

Images