Network alignment method, device and equipment based on local structural feature enhancement

Abstract

The present application relates to a network alignment method, device and apparatus based on local structural feature enhancement. The method comprises: inputting two network data sets into a network alignment model for training until a trained network alignment model with the ability to align two networks is obtained, the network alignment model comprises an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set to alleviate the noise in the original network structure, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure to increase local features and ultimately improve the accuracy of network alignment.

Description

Technical Field

[0001] The present application relates to the technical field of network alignment, and in particular to a network alignment method, device and equipment based on local structural feature enhancement. Background Art

[0002] In today's world, networks are almost everywhere. From computer vision, bioinformatics, network mining, chemistry to social network analysis, various networks naturally emerge in many high-impact application fields. Typically, network alignment (i.e., finding the correspondence of nodes in different networks) is actually the first step in any data mining task in these applications. For example, by linking users of different social networking sites, products from one website (such as Taobao) can be recommended to users of another website (such as WeChat). In bioinformatics, integrating different tissue-specific protein-protein interaction (PPI) networks has led to significant improvements in candidate gene prioritization.

[0003] Generally speaking, assuming that there is an inherent correspondence between two (or more) nodes, the purpose of network alignment is to find a set of node pairs in two (or more) networks. Considering that the scale of networks in real life is large, and the network alignment model based on spectral methods is not computationally efficient, network alignment methods based on network representation learning have emerged. Although methods based on network representation learning have also been proven to be effective in learning the representation of two networks, current network alignment methods based on network representation learning, such as PALE and IONE, ignore the noise problem of the network itself. The network alignment method based on iterative deep graph learning ignores the problem of data sparsity. At the same time, the model pays more attention to global structural features and ignores local structural features, which makes the model less effective on sparse data sets. Summary of the invention

[0004] Based on this, it is necessary to provide a network alignment method, device and equipment based on local structural feature enhancement, which can enhance local features and alleviate network noise problems in response to the above technical problems.

[0005] A network alignment method based on local structural feature enhancement, the method comprising:

[0006] Obtaining a training set, wherein the training set includes two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0007] Inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0008] Two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0009] In one embodiment, when the decoding part adopts the knowledge representation learning method, the scoring function between each node in one optimized network structure and each node in another optimized network structure is calculated respectively, and whether the two nodes are aligned is predicted based on the scoring function.

[0010] In one embodiment, the scoring function is calculated based on two nodes and an edge between the two nodes.

[0011] In one embodiment, in the decoding part, a scoring function corresponding to any one of the TransE, TransH, DistMult, ComplEx and RotatE models is selected based on a knowledge representation learning method to perform network alignment prediction.

[0012] In one embodiment, when training the network alignment model, the encoding part and the decoding part are trained using loss functions respectively, wherein, in the decoding part, the loss function corresponding to the selected model is used as the prediction loss function to train the decoding part, and in the encoding part, the mixed loss function of the graph regularization loss function and the prediction loss function is used to train the encoding part.

[0013] In one embodiment, when training the network alignment model, a Bootstrapping algorithm is used to add newly generated node pairs to the training set to train the network alignment model.

[0014] A network alignment device based on local structural feature enhancement, the device comprising:

[0015] A training set acquisition module, used to acquire a training set, wherein the training set includes two network data sets, each of which includes a collection of multiple nodes and a collection of edges representing the relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0016] A network alignment model training module, used for inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0017] The network alignment module is used to obtain two network data sets to be aligned, and input the two network data sets into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0018] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0019] Obtaining a training set, wherein the training set includes two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0020] Inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0021] Two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0022] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0023] Obtaining a training set, wherein the training set includes two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0024] Inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0025] Two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0026] The above-mentioned network alignment method, device and equipment based on local structural feature enhancement, by adopting an iterative deep graph learning method in the network alignment model to iteratively optimize the input network structure to obtain an updated network structure that is more suitable for network alignment, which can alleviate the noise existing in the network structure, and then adopt a knowledge representation learning method to perform network alignment based on the updated network structure to enhance the local features of the network structure, so as to improve the accuracy of network alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of a network alignment method based on local structural feature enhancement in one embodiment;

[0028] Figure 2 A schematic diagram of the structure of a network alignment model in one embodiment;

[0029] Figure 3 This is a diagram showing how Success@10 changes with the proportion of the training set in the experiment;

[0030] Figure 4 This is a schematic diagram of the change of MAP with the proportion of training sets in the experiment;

[0031] Figure 5 It is a structural block diagram of a network alignment device based on local structural feature enhancement in one embodiment;

[0032] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0034] like Figure 1 As shown, a network alignment method based on local structure feature enhancement is provided, comprising the following steps:

[0035] Step S100, obtaining a training set, the training set including two network data sets, each of which includes a collection of multiple nodes and a collection of edges representing the relationship between two nodes, and two nodes with matching relationships in the two network data sets have an anchor link between them and form a node pair;

[0036] Step S110, inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0037] Step S120, obtaining two network data sets to be aligned, and inputting the two network data sets into a trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0038] In this embodiment, steps S100 and S110 are steps of training a network alignment model, and S120 is applying the trained network alignment model. In this article, steps S100 and S110 are mainly described.

[0039] In step S110, the two network data sets in the training set represent the network structure as G = (V, E), where V and E represent the node set and edge set of G respectively, and |V| and |E| represent the number of nodes and edges. i and u j , if there is a link between them, then there is an edge e i exists in the edge set.

[0040] As an example, taking social networks as an example, users can be regarded as nodes, and the relationship between observed users as edges. If there is no edge between two users, it means that the two users are not friends (or they are not observed). Users may usually register on two social networks, and some friends on these social networks may overlap. Therefore, finding the corresponding relationship between nodes in two networks is the network alignment task.

[0041] For network alignment, that is, given two networks (G s ,G t ), G s represents the source network, G t represents the target network. The goal of network alignment is to find a mapping π:V s →V t , find the sTo V t One-to-one correspondence, that is, v t =π(u s ), similarly, u s =π -1 (v t ), where u s and v t Respectively represent G s and G t Nodes in .

[0042] like Figure 2 As shown in the figure, the network alignment model includes an encoding part and a decoding part, wherein the encoding part is a neural network based on iterative deep graph learning to optimize and iterate the input network structure to obtain an updated network structure, and this network structure is more suitable for downstream tasks, namely network alignment. Iterative deep graph learning can effectively alleviate the noise in the original network structure. The decoding part is a neural network based on knowledge representation learning, which can enhance the local structural features of the network structure for network alignment prediction. In the following, the encoding part based on iterative deep graph learning is first introduced, and then the decoding part of knowledge representation learning is introduced.

[0043] Specifically, iterative deep graph learning is a new iterative method that aims to search for an implicit graph structure to enhance the initial graph structure of downstream prediction tasks. Based on the proposed stopping criterion, the iterative method dynamically stops in each mini-batch when the learned graph structure is close to the optimized graph (relative to the downstream task). In addition, the graph construction process can be optimized for downstream tasks (alignment tasks) in an end-to-end manner. In the method, various network structures can be represented in the form of graph structures, so the graph structure can be updated using iterative deep graph learning methods.

[0044] In previous methods, the graph learning problem is modeled as learning a joint discrete probability distribution over the edges of the graph, and these methods have shown good performance. However, since they optimize edge connectivity by assuming that the graph nodes are known, they cannot handle the inductive setting (using new nodes during testing). To overcome this problem, in this embodiment, the graph structure learning problem is converted into similarity metric learning.

[0045] Specifically, common methods for learning graph similarity metrics include cosine similarity, radial basis function (RBF) kernel, and attention mechanism. A good similarity metric function should be learnable and expressive. Although the iterative deep graph learning framework is agnostic to various similarity metric functions, without losing generality, in this embodiment, weighted cosine similarity is used as the metric function, that is, s ij =cos(w⊙v i ,w⊙v j ), where ⊙ represents the Hadamard product, w is a learnable weight vector with the same dimension as the input vector v i and v j The same, and learn different dimensions of the highlighting vectors. It is worth noting that the two input vectors can be either raw node features or computed node embeddings.

[0046] At the same time, in order to stabilize the learning process and improve the expressive power, the similarity measurement function is extended to a multi-head version. Specifically, m weight vectors are used, each weight vector represents a perspective, and the above similarity function is used to calculate m independent similarity matrices, and their average value is used as the final similarity. The formula is as follows:

[0047]

[0048] In formula (1), Calculate the two input vectors v for the pth perspective i and v j The cosine similarity between , where each view considers a part of the semantics captured in the vector.

[0049] Normally, the adjacency matrix calculated by the similarity measure should be non-negative, but s ij The range is between [-1, 1]. In addition, many underlying graph structures are much sparser than fully connected graphs, which are not only computationally expensive but may also introduce noise (i.e., unimportant edges). Therefore, by only considering the ε neighborhood of each node, we continue to extract the symmetric sparse non-negative adjacency matrix A from S. Specifically, we mask out the elements in S that are less than the non-negative threshold ε and set the value to 0.

[0050] Although the initial graph may be noisy, it usually still contains rich and useful information about the true graph topology. Ideally, the learned graph structure A can complement the original graph topology A. (0), to formulate an optimized graph for GNNs for the alignment task. Therefore, under the mild assumption that the optimized graph structure may be a "transfer" of the initial graph structure, the learned graph is combined with the initial graph as follows:

[0051]

[0052] In formula (2), I (0) =D (0)-1 / 2 A (0) D (0)-1 / 2 is the normalized adjacency matrix of the initial graph. A (t) and A (1) They are the two adjacency matrices of the tth and first iterations calculated by formula 1. Further normalizing the adjacency matrix gives f(A) ij =A ij / ∑ j A ij .

[0053] It is worth noting that A (0) is calculated based on the original node feature X, while A (t) is based on the previously updated node embedding Z (t-1) The node is embedded in Z (t-1) is optimized for the alignment prediction task. Therefore, the final learned graph structure is taken as their linear combination and weighted by the hyperparameter η to combine the advantages of both. Finally, another hyperparameter λ is used to balance the balance between the learned graph structure and the initial graph structure.

[0054] At the same time, the graph learning framework is independent of various GNN architectures (which take node feature matrices and adjacency matrices as input to calculate node embeddings) and prediction tasks. In the iterative deep graph learning network in this embodiment, two layers of GCN are used, where the first layer (denoted as GNN 1 ) maps the original node feature X to the intermediate embedding space, and the second layer (represented as GNN 2 ) further maps the intermediate node embedding Z to the output space.

[0055] At the same time, the proposed graph learning framework is independent of various GNN architectures (taking node feature matrix and adjacency matrix as input to calculate node embedding) and prediction tasks. In this paper, a two-layer GCN is adopted, in which the first layer (denoted as GNN 1 ) maps the original node feature X to the intermediate embedding space, and the second layer (represented as GNN 2 ) further maps the intermediate node embedding Z to the output space.

[0056]

[0057] In formula (3), σ(·) and l(·) are the task-related output function and loss function, respectively. In the aligned prediction task, σ(.) is the softmax function and l(.) is the cross entropy function used to calculate the prediction loss. MP(·, ·) is a message passing function, and in GCN, The feature / embedding matrix F and the normalized adjacency matrix are used (obtained using formula 2).

[0058] Although we will learn Figure A (t) Compared with the initial figure A (0) Combining is an effective way to approximate the optimization graph, but learning graph A (t) The quality of the final image The quality of plays an important role. In practice, controlling the generated learning graph A (t) The smoothness, connectivity, and sparsity of X are very important, which faithfully reflects the graph topology with respect to the initial node attributes X and downstream tasks.

[0059] Each column of the feature matrix X is considered as a graph signal. A widely adopted assumption of graph signals is that the values ​​vary smoothly between adjacent nodes. Given an undirected graph with a symmetric weighted adjacency matrix A, a set of n graph signals It is usually measured by Dirichlet energy, and the formula is as follows:

[0060]

[0061] In formula (4), tr(·) represents the trace of the matrix, L = DA is the graph Laplacian operator, and D = ∑ j A ij It can be seen that minimizing Ω(A, X) can force adjacent nodes to have similar characteristics, thereby enhancing the smoothness of the graph signal on the graph associated with A.

[0062] However, minimizing only the smoothness loss will lead to a zero solution A = 0. In addition, it is also desirable to be able to control the sparsity of the generated graph. Therefore, in the iterative deep graph learning network, additional constraints are imposed on the learned graph, as shown below:

[0063]

[0064] In formula (5), ||·|| F represents the Frobenius norm of the matrix. The first term penalizes the formation of disconnected graphs by a logarithmic barrier, and the second term controls sparsity by penalizing the large degree of the first term.

[0065] Then, the regularization loss of the overall graph is defined as the above loss It is the sum of , which can control the smoothness, connectivity and sparsity of the learning graph, where α, β and χ are all non-negative hyperparameters.

[0066] When training this part of the neural network, compared with previous work that directly optimizes the adjacency matrix based on graph regularization loss or task-related prediction loss, in the iterative deep graph learning network, a hybrid loss function that combines task prediction loss and graph regularization loss is minimized, that is, To jointly and iteratively learn the graph structure and GNN parameters. The task prediction loss is the loss function of the decoding part of the neural network, which will be explained later.

[0067] After the encoding part iteratively updates the network structure, the decoding part performs alignment prediction on the network. In this part, the knowledge representation learning method is adopted. In this method, the network structure is represented by multiple nodes and edges between two nodes with relationships. In knowledge representation learning, a knowledge graph can be represented as Among them, and denote the set of entities, relations, and triples, respectively. The head entity can be denoted as h, the relation as r, and the tail entity as t. Thus, (h, r, t) denotes a triple, and the bold letters h, r, t denote the embedding of (h, r, t). The number of entities, relations, and embedding dimensions are denoted by n. e 、n r and d. The nodes in the network can be regarded as entities and the edges can be regarded as relationships, so the knowledge representation learning method can be applied to the network alignment task.

[0068] Specifically, because the nodes in the network can be regarded as entities in knowledge representation learning, and the edges in the network can be regarded as relations, the alignment problem in the network can be regarded as the link prediction problem in knowledge representation learning.

[0069] Furthermore, when the knowledge representation learning method is used in the decoding part, the scoring function between each node in one optimized network structure and each node in another optimized network structure is calculated, and whether the two nodes are aligned is predicted based on the scoring function. The scoring function is calculated based on the two nodes and the edges between the two nodes.

[0070] In this embodiment, in the decoding part, any one of the TransE, TransH, DistMult, ComplEx and RotatE models can be selected from the knowledge representation learning method to adapt to the network alignment framework, so as to be used as a method for alignment prediction. Each model has a corresponding scoring function. The following is an introduction to the scoring function of each model:

[0071] In the TransE model, the idea is to expect that the embedding of the nodes in the source network can be close enough to the embedding of the corresponding nodes in the target network after the translation of the edge embedding. Therefore, the score function of the TransE model can be expressed by the following formula:

[0072] f TransE (u+e,v)=||u+ev|| (6)

[0073] In formula (6), u and v represent the embedding of nodes, e represents the embedding of edges,

[0074] For the TransH model, it is to overcome the defects of TransE in edge modeling, so that nodes have distributed representations when involving different edges. For an edge, the model translates the edge-specific translation vector d e The hyperplane w located on a specific edge e Instead of in the node embedding space, its scoring function is expressed as follows:

[0075]

[0076] In formula (7), u and v represent the embedding of nodes, e represents the embedding of edges, .

[0077] The DistMult model uses bilinear encoding, in which the embedding of nodes and edges can be learned by a neural network. The first layer projects a pair of input nodes onto a low-dimensional vector, and the second layer combines the two vectors into a scalar that is compared by a scoring function with edge-specific parameters. The scoring function is:

[0078] f DistMult (u, v) = u T B e v (8)

[0079] In formula (8), u and v represent the embedding of the node, B e is an edge-specific parameter.

[0080] The ComlpEx model introduces complex vector space into embedding, and its scoring function is expressed by the following formula:

[0081]

[0082] In formula (9), Re(·) represents the real part, and Im(·) represents the imaginary part. u and v represent the embedding of nodes, and e represents the embedding of edges. Indicates v kconjugate. Θ corresponds to the embedding

[0083] The RotatE model, like the ComplEx model, models nodes and edges in a complex vector space. The difference is that RotatE limits the modulus of the edge vector to 1, making it a rotation vector from the source network node to the corresponding node of the target network. Therefore, its scoring function is expressed as:

[0084]

[0085] In formula (10), is the Hadmard product, u and v represent the embedding of nodes, e represents the embedding of edges, e i express Each element of The modulus is limited to |e i |=1.

[0086] In this way, there is greater flexibility in selecting the model of the decoding part, and the model of the decoding part can be selected according to the specific situation.

[0087] In this embodiment, the network alignment model is actually a joint model composed of a neural network based on an iterative deep graph learning method and a neural network based on knowledge representation learning. Figure 2 In the paper, the specific name of the knowledge representation learning model is not written in the framework diagram, because any knowledge representation learning model can be selected to achieve network alignment.

[0088] When training the network alignment model, the loss function of the neural network based on the iterative deep graph learning method has been mentioned in the previous article. Next, the loss function of the neural network based on knowledge representation learning is introduced:

[0089] When the knowledge representation learning method performs alignment prediction, different loss functions need to be trained for different models.

[0090] For the TransE model, in order to learn translation-based embeddings, a margin-based ranking criterion is minimized on the training set:

[0091]

[0092] In formula (11), [x] + =max{0,x},S′ (u，v) Denotes the negative node alignment set constructed by reorganizing (u, v), that is, G s or G tA randomly selected node replaces u or v in . u and v as well as u′ and v′ represent the embedding of nodes, and e represents the embedding of edges. 1 >0 is the boundary hyperparameter separating positive node alignment and negative node alignment in the TransE model.

[0093] For the TransH model, in order to encourage the distinction between correctly aligned nodes and incorrectly aligned nodes, the following margin-based ranking loss is used:

[0094]

[0095] In formula (12), [x] + =max{0,x},S′ (u，v) Denotes the negative node alignment set constructed by reorganizing (u, v), that is, G s or G t A randomly selected node replaces u or v in . u and v as well as u′ and v′ represent the embedding of the node. δ 2 >0 is the boundary hyperparameter separating positive node alignment and negative node alignment in the TransH model.

[0096] In the DistMult model, it can be learned by minimizing a margin-based ranking objective that encourages positive node alignments to have higher scores than any negative node alignments. The goal of model training is to minimize the margin-based ranking loss:

[0097]

[0098] In formula (13), [x] + =max{0,x},S′ (u，v) Denotes the negative node alignment set constructed by reorganizing (u, v), that is, G s or G t A randomly selected node replaces u or v in . u and v as well as u′ and v′ represent the embeddings of the nodes.

[0099] In the ComplEx model, the negative log-likelihood of the logistic model is minimized by regularizing the parameters Θ of the considered model. The model is trained using mini-batch stochastic gradient descent and AdaGrad to adjust the learning rate:

[0100]

[0101] In formula (14), when the node pair is a correct aligned node, Y euv = 1, when the node pair is an incorrectly aligned node, Y euv= -1. λ can be verified in {0.1, 0.03, 0.01, 0.003, 0.001, 0.0003, 0.00001, 0.0}.

[0102] The loss function of the RotatE model is somewhat different from the above model. It adopts the negative sampling loss function of self-adversarial training for training:

[0103]

[0104] In formula (15), δ 3 is the boundary hyperparameter, σ is the sigmoid function, (u′ i , e, v′ i ) is the i-th negative alignment node. Specifically, p(v′ i , e, v′ i ) can be defined as:

[0105]

[0106] In formula (16), α represents the sampling weight.

[0107] The loss functions of the above models will also be used as the loss functions of the iterative deep graph learning neural network. The loss function part iteratively learns a better network structure, which is the task prediction loss mentioned above.

[0108] That is to say, when training the network alignment model, the encoding part and the decoding part are trained using loss functions respectively, wherein, in the decoding part, the loss function corresponding to the selected model is used as the prediction loss function to train the decoding part, and in the encoding part, a mixed loss function of the graph regularization loss function and the prediction loss function is used to train the encoding part.

[0109] In this embodiment, when the network alignment model is trained, the Bootstrapping algorithm is used to add the newly generated node pairs into the training set to train the network alignment model.

[0110] Specifically, when training the network alignment model, especially when training the decoding part, it is necessary to apply it to the node pairs that already have anchor chains in the training set, so that the parameters of this part can be adjusted through the loss function until the loss function meets the preset requirements. However, not all matching relationships in the training set are set with anchor chains, so new node pairs will appear during training, and these new node pairs will be added to the training set to train the model again, which effectively alleviates the problem of sparse training data. Similarly, after the model is initially trained, it will also be tested using the model in the test set to adjust the accuracy of the model.

[0111] In this paper, this method is also verified by experiments, including:

[0112] The experiment was conducted on 1 real-world dataset (2 real-world networks), and only the datasets Flickr and Myspace were selected for the experiment. The reason is that in this dataset, the number of edges of Flickr and Myspace is relatively small compared to the number of nodes, the network structure is relatively sparse, and the number of attributes is also small. The purpose of the model proposed in this paper is to further learn better node representations and local network structure representations to improve the effect of the model. Therefore, compared with the datasets Allmovie and Imdb and the datasets Douban Online and Douban Offline, the datasets Flickr and Myspace can better reflect the effect of the model. The relevant information of the dataset is shown in Table 1.

[0113] Flickr and Myspace datasets: The two subnets of Flickr and Myspace are collected in the article and then processed as described in the article. The Flickr subnet contains 6714 nodes and the Myspace subnet contains 10733 nodes. The gender of the user is used to represent the node attribute, and only partial groundtruth is available for alignment.

[0114] Table 1 Data information of real-world networks

[0115]

[0116] The relevant settings in this experiment are:

[0117] Among them, the setting of the measurement indicators uses the most advanced indicators to evaluate the performance of network alignment from the prediction perspective and the ranking perspective. From the prediction perspective, Success@q is used, which indicates whether a true positive match appears in the first q candidates. More specifically, for each anchor pair In the ground truth, if the alignment score is In the alignment matrix S The node is within the range of the first q highest values ​​in the row. The alignment output is recorded as a success case:

[0118]

[0119]

[0120] In formulas (17) and (18), Success B / A @q represents the number of hits of each node in network B / A in the test set in network A / B, and Success@q is the average of the two-way counts.

[0121] From the ranking perspective, the average precision MAP (ie, MRR in the paired setting) is used as the measurement indicator:

[0122]

[0123]

[0124] In formulas (19) and (20), rank i It refers to the ranking position of the i-th anchor node in the test set in network A / B with respect to its ground truth counterpart in network B / A. Like Success@q, MAP is also the average of the two-way counts.

[0125] In the experiment, the value of q in Success@q is 1, 3, 5 and 10. A good network alignment model should have high Success@q value and MAP value.

[0126] In the experiment, a network alignment method based on iterative deep graph learning was selected as the baseline of the model. The network alignment model with different knowledge representation learning methods was compared with the baseline model.

[0127] As for the hyperparameters in the network alignment model, the embedding dimension d is set to 100, and the boundary hyperparameter δ is set to 100 in the loss function of the alignment prediction. 1 , δ 2 , δ 3 Set to 1.

[0128] The results are averaged over 10 runs to mitigate randomness. All experiments are performed on a GeForce RTX2080Ti server with a processor i7-9700K and 16GB of memory and implemented with PyTorch.

[0129] In order to verify the effect of the network alignment model on the network alignment task, in this experiment, the model is compared with the network alignment method based on iterative deep graph learning on the datasets Flickr and Myspace. The ratio of the training set to the test set is 0.8:0.2. The experimental results are shown in Table 2. It can be seen from the experimental results that when the alignment prediction part is set to RotatE, the model has the most outstanding effect, especially in terms of Success@1 and MAP indicators. However, other knowledge representation models did not improve the effect of the model after replacing the decoding part. The reason may be that in the network dataset, the edges in the network are not given as much information as the relationships in the knowledge graph. They are simply set to exist or not exist, which greatly reduces the effect of the knowledge representation model in alignment. In RotatE, the modulus of the edge is limited to 1, and it is used as the rotation from the source network node to the target network node. The edge only plays the role of direction transformation, so the effect of the model will be more prominent. At the same time, RotatE uses the negative sampling loss function of self-adversarial training. This loss function is different from other methods. The improvement of the model effect may also be related to the training mode.

[0130] Table 3 further compares the model performance when the ratio of training set to test set is 0.2:0.8. It can be seen that under weak supervision, the model has the best performance when the encoding part is set to ComplEx, especially in terms of Success@1, Success@3 and MAP indicators. When the encoding part is set to TransH, the results of Success@5 and Success@10 are better. This shows that under weak supervision, the setting of hyperplane and complex vector space can improve the performance of the model.

[0131] Table 2 Experimental results of the model on Flickr and Myspace datasets (the ratio of training set to test set is 0.8:0.2)

[0132]

[0133] Table 3 Experimental results of the model on Flickr and Myspace datasets (the ratio of training set to test set is 0.2:0.8)

[0134]

[0135] In order to further understand the impact of the ratio of training set to test set in the model on the experimental results, this parameter was tested in this experiment. During the experiment, only the value of this parameter was changed, and other parameters were kept unchanged. During the experiment, the ratio of the training set increased from 0.1 to 0.9, increasing by 0.1 each time. The experimental results are as follows: Figure 3 and 4shown.

[0136] Figure 3 The Success@10 indicator changes with the training set ratio. It can be seen from the figure that when the training set ratio is small, that is, when the training set ratio is 0.1, 0.2 and 0.3, the encoding part is better when the knowledge representation learning model is used. This shows that under weak supervision, the knowledge representation learning model works better as the decoding part. When the training set ratio increases, the RotatE model as the encoding part has a similar effect to the original network alignment model based on iterative deep graph learning, while the advantages of the TransH and ComplEx models as the encoding part gradually become less obvious.

[0137] and Figure 4 This is the change of the indicator MAP with the proportion of the training set. From the overall trend, it is similar to the change of the Success@10 indicator. When the encoding part is a knowledge representation learning model, when the training set proportion is 0.1, 0.2, and 0.3, the model effect is better than the network alignment model based on iterative deep graph learning. However, in the entire change process, the fluctuations of the MAP indicator changes of each model are greater. Unlike other models, when the RotatE model is used as the encoding part, the effect is significantly improved when the training set proportion changes from 0.7 to 0.8, but when it changes from 0.8 to 0.9, the model effect is not significantly improved. On the other hand, compared with the TransE, TransH and ComplEx models, when the DistMult model is used as the encoding part, the model effect improves more significantly in the MAP indicator as the training set proportion increases, indicating the effectiveness of bilinear coding in alignment tasks.

[0138] Summary: The network alignment model based on the local structural feature enhancement proposed in this method takes into account that the network alignment method based on iterative deep graph learning ignores the noise in the node features and does not consider the problem of data sparsity. It focuses on the role of knowledge representation learning in node feature learning and alleviating the problem of data sparsity, and adds the knowledge representation learning model to the decoding part of the network alignment model based on iterative deep graph learning. The model based on local structural feature enhancement tries 5 classic knowledge representation learning models for alignment mode, and adds the Bootstrapping algorithm during the training process to solve the noise problem and data sparsity problem in network alignment. The model selects sparse data sets Flickr and Myspace for experiments. The experimental results show that when the proportion of training sets is small, the effect of the network alignment method based on local structural feature enhancement is superior to the network alignment method based on iterative deep graph learning. When the proportion of training sets increases, the RotatE model still has a good effect when used as the encoding part.

[0139] In the above-mentioned network alignment method based on local structural feature enhancement, the characterization of local structural features is enhanced by the knowledge representation learning method, while alleviating the problems of network noise and data sparsity. In the model training, a Bootstrapping algorithm is designed to input the newly generated alignment node pairs in the training into the training set for further training, which can further alleviate the problem of data sparsity.

[0140] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0141] In one embodiment, Figure 5 As shown, a network alignment device based on local structural feature enhancement is provided, comprising: a training set acquisition module 200, a network alignment model training module 210 and a network alignment module 220, wherein:

[0142] The training set acquisition module 200 is used to acquire a training set, wherein the training set includes two network data sets, each of which includes a collection of multiple nodes and a collection of edges representing the relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0143] A network alignment model training module 210 is used to input the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0144] The network alignment module 220 is used to obtain two network data sets to be aligned, and input the two network data sets into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0145] For the specific definition of the network alignment device based on local structural feature enhancement, please refer to the definition of the network alignment method based on local structural feature enhancement above, which will not be repeated here. Each module in the above-mentioned network alignment device based on local structural feature enhancement can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0146] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a network alignment method based on local structural feature enhancement is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0147] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0148] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0149] Obtaining a training set, wherein the training set includes two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0150] Inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0151] Two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0152] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0153] Obtaining a training set, wherein the training set includes two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair;

[0154] Inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure;

[0155] Two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

[0156] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0157] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A network alignment method based on local structural feature enhancement, characterized in that: The method comprises: Select two network data sets for network alignment, the two network data sets are data sets of two real-world social networks, the two social network data sets are Flickr and Myspace data sets respectively; in the social networks, users are regarded as nodes, and the relationship between users is regarded as edges; Acquire a training set, wherein the training set includes the two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing a relationship between two nodes, and two nodes in a matching relationship in the two network data sets have an anchor link between them and form a node pair; Input the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure. When the decoding part uses the knowledge representation learning method, a scoring function between each node in one of the optimized network structures and each node in the other optimized network structure is calculated respectively, and whether the two nodes are aligned is predicted according to the scoring function; The two network data sets to be aligned are obtained, and the two network data sets are input into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

2. The network alignment method according to claim 1, characterized in that: The scoring function is calculated based on two nodes and the edge between the two nodes.

3. The network alignment method according to claim 1, characterized in that: In the decoding part, based on the knowledge representation learning method, a scoring function corresponding to any one of the TransE, TransH, DistMult, ComplEx and RotatE models is selected to perform network alignment prediction.

4. The network alignment method according to claim 3, characterized in that: When training the network alignment model, the encoding part and the decoding part are trained using loss functions respectively, wherein, in the decoding part, the loss function corresponding to the selected model is used as the prediction loss function to train the decoding part, and in the encoding part, the mixed loss function of the graph regularization loss function and the prediction loss function is used to train the encoding part.

5. The network alignment method according to claim 4, characterized in that: When training the network alignment model, a Bootstrapping algorithm is used to add newly generated node pairs into the training set to train the network alignment model.

6. A network alignment device based on local structural feature enhancement, characterized in that: The device comprises: A network dataset selection module is used to select two network datasets for network alignment, wherein the two network datasets are datasets of two real-world social networks, and the two social network datasets are respectively a Flickr dataset and a Myspace dataset; in the social networks, users are regarded as nodes, and the relationships between users are regarded as edges; A training set acquisition module, used to acquire a training set, wherein the training set includes the two network data sets, each of the network data sets includes a collection of multiple nodes and a collection of edges representing the relationship between two nodes, and two nodes with matching relationships in the two network data sets have an anchor link between them and form a node pair; A network alignment model training module, used for inputting the training set into the network alignment model for training until a trained network alignment model capable of aligning two networks is obtained, wherein the network alignment model includes an encoding part and a decoding part, the encoding part uses an iterative deep graph learning method to optimize the network structure represented by the network data set, and the decoding part uses a knowledge representation learning method to perform alignment prediction on the optimized network structure. When the decoding part uses the knowledge representation learning method, a scoring function between each node in one of the optimized network structures and each node in the other optimized network structure is calculated respectively, and whether the two nodes are aligned is predicted according to the scoring function; The network alignment module is used to obtain the two network data sets to be aligned, and input the two network data sets into the trained network alignment model to obtain node pairs with matching relationships in the two networks.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

Images