Network alignment method and apparatus
View Patent ↗A network alignment method comprises the steps of: receiving two networks as inputs and performing a neural network operation respectively, vectorizing a plurality of nodes of each of the two networks; calculating a dual-perception similarity based on an embedding similarity between the vectorized nodes of each of the two networks, and a Tversky similarity representing a ratio of the number of previously aligned nodes included in a neighboring node to the normalized number of neighboring nodes of each combined node when configuring a node pair by combining nodes that are not aligned in the two networks; and selecting node pairs to be aligned among a plurality of nodes of the two networks based on the dual-perception similarity, thereby partially aligning the two networks so that the number of node pairs aligned in the two networks gradually increases according to the dual-perception similarity updated according to the two partially aligned networks.
1. A network alignment method performed by a computing device having one or more processors and a memory storing one or more programs executed by the one or more processors, the method comprising the steps of:
receiving a source network and a target network as inputs and performing a neural network operation respectively, thereby vectorizing a plurality of nodes of each of the two networks;
calculating a dual-perception similarity based on an embedding similarity, which is a similarity between the vectorized nodes of each of the two networks, and a Tversky similarity representing a ratio of the number of previously aligned nodes included in a neighboring node to the normalized number of neighboring nodes of each combined node when configuring a node pair by combining nodes that are not aligned in the two networks; and
selecting node pairs to be aligned among a plurality of nodes of the two networks based on the dual-perception similarity, thereby partially aligning the two networks, and iteratively partially aligning so that the number of node pairs aligned in the two networks gradually increases according to the dual-perception similarity updated according to the two partially aligned networks,
wherein the step of vectorizing includes
performing a neural network operation on the two networks with two neural networks having the same structure having L layers and the same learning weight, to obtain L source embedding vector sets (H s (l) ) and L target embedding vector sets (H t (l) ) output from the L layers of the two neural networks,
wherein the step of calculating a dual-perception similarity includes
obtaining the embedding similarity by weighting a source embedding vector set (H s (l) ) with a target embedding vector set (H t (l) ) output from the same layer (l) of the L source embedding vector sets and the L target embedding vector sets,
configuring a plurality of mock node pairs by combinations of the remaining nodes except for the previously aligned node pairs in two networks that are repeatedly partially aligned, checking neighboring nodes of the nodes combined in the mock node pairs and nodes of other networks aligned with the previously aligned nodes among the neighboring nodes, thereby iteratively calculating the Tversky similarity, and
iteratively calculating the dual-perception similarity by element-multiplying the embedding similarity and the iteratively calculated Tversky similarity.
2. The network alignment method according to claim 1 , wherein the embedding similarity is calculated according to the equation
S
emb
=
∑
l
H
s
(
l
)
H
t
(
l
)
⊤
wherein denotes a transpose matrix of the target embedding vector set (H t (l) ) output from the l-th layer (l).
3. The network alignment method according to claim 1 ,
wherein the step of iteratively calculating the Tversky similarity includes
configuring a mock node pair (u, v) by combining nodes that are not previously aligned in the two networks,
obtaining a set of neighboring nodes ( , ) of each node of the mock node pair (u, v),
searching previously aligned nodes among the nodes included in the neighbor node set ( ) obtained from the source network among the two networks,
checking neighbor aligned cross-network nodes that are nodes of the target network among the two networks corresponding to the previously aligned nodes, thereby calculating the Tversky similarity.
4. The network alignment method according to claim 1 ,
wherein the step of iteratively partially aligning includes partially aligning by selecting a predetermined number of node pairs having the highest dual-perception similarity in each iteration.
5. The network alignment method according to claim 1 , wherein the step of iteratively partially aligning includes partially aligning by selecting node pairs whose dual-perception similarity is equal to or greater than the specified criterion similarity in each iteration.
6. A network alignment apparatus having one or more processors and a memory storing one or more programs executed by the one or more processors,
wherein the processors
receive a source network and a target network as inputs and perform a neural network operation respectively, thereby vectorizing a plurality of nodes of each of the two networks,
calculate a dual-perception similarity based on an embedding similarity, which is a similarity between the vectorized nodes of each of the two networks, and a Tversky similarity representing a ratio of the number of previously aligned nodes included in a neighboring node to the normalized number of neighboring nodes of each combined node when configuring a node pair by combining nodes that are not aligned in the two networks, and
select node pairs to be aligned among a plurality of nodes of the two networks based on the dual-perception similarity, thereby partially aligning the two networks, and iteratively partially align so that the number of node pairs aligned in the two networks gradually increases according to the dual-perception similarity updated according to the two partially aligned networks,
wherein the processors perform a neural network operation on the two networks with two neural networks having the same structure having L layers and the same learning weight, to obtain L source embedding vector sets and L target embedding vector sets output from the L layers of the two neural networks,
wherein the processors
obtain the embedding similarity by weighting a source embedding vector set (H s (l) ) with a target embedding vector set (H t (l) ) output from the same layer (l) of the L source embedding vector sets and the L target embedding vector sets,
configure a plurality of mock node pairs by combinations of the remaining nodes except for the previously aligned node pairs in two networks that are iteratively partially aligned, check neighboring nodes of the nodes combined in the mock node pairs and nodes of other networks aligned with the previously aligned nodes among the neighboring nodes, thereby iteratively calculating the Tversky similarity, and
iteratively calculate the dual-perception similarity by element-multiplying the embedding similarity and the iteratively calculated Tversky similarity.
7. The network alignment apparatus according to claim 6 ,
wherein the processors calculate the embedding similarity according to the equation
S
emb
=
∑
l
H
s
(
l
)
H
t
(
l
)
⊤
wherein denotes a transpose matrix of the target embedding vector set (H t (l) ) output from the l-th layer (l).
8. The network alignment apparatus according to claim 6 ,
wherein, in order to calculate the Tversky similarity, the processors configure a mock node pair (u, v) by combining nodes that are not previously aligned in the two networks,
obtain a set of neighboring nodes ( , ) of each node of the mock node pair (u, v),
search previously aligned nodes among the nodes included in the neighbor node set ( ) obtained from the source network among the two networks, and
check neighbor aligned cross-network nodes that are nodes of the target network among the two networks corresponding to the previously aligned nodes, thereby calculating the Tversky similarity.
9. The network alignment apparatus according to claim 6 ,
wherein the processors select and partially align a predetermined number of node pairs having the highest dual-perception similarity in each iteration.
10. The network alignment apparatus according to claim 6 ,
wherein the processors select and partially align node pairs whose dual-perception similarity is equal to or greater than the specified criterion similarity in each iteration.