An unsupervised network alignment method and device for cross-platform user identification

Abstract

This invention relates to an unsupervised network alignment method for cross-platform user identification. This method uses a set of high-confidence pseudo-anchor pairs as a shared structural reference across the network, encoding each node with its topological distance to all pseudo-anchors in the network. This maps nodes from different networks to the same reference, allowing for direct comparison. Since the node encoding is directly calculated from the network topology without hop-by-hop message propagation, the structural coordinates of nodes far from pseudo-anchors change more smoothly with topological distance. Furthermore, this invention establishes cluster-level correspondence across networks, performing pseudo-anchor filtering only within corresponding clusters. This effectively narrows the search range, reduces interference from irrelevant node pairs, and improves the reliability of pseudo-anchor filtering, thereby enhancing the accuracy of cross-platform user identification tasks.

Description

Technical Field

[0001] This invention belongs to the field of network alignment technology, and specifically relates to an unsupervised network alignment method and device for cross-platform user identification. Background Technology

[0002] Network alignment is a key technology in graph data analysis and artificial intelligence. Networks are commonly used modeling tools for complex systems such as social relationships, academic collaborations, and biomolecular interactions. Networks from different sources often contain overlapping entities, and network alignment establishes node correspondences by comparing the consistency of the structure and attributes of nodes in two networks. Cross-platform user identification is an important application scenario for network alignment. By associating the accounts of the same user on different social platforms, applications such as user profile integration, abnormal account identification, and personalized recommendations can be achieved. In practical applications, due to factors such as privacy protection, limited data sharing, and high costs of manual annotation, it is often difficult to obtain a large number of known corresponding nodes. Therefore, unsupervised network alignment has become an important research direction.

[0003] To establish correspondences between nodes, most existing network alignment methods rely on high-confidence anchor pairs as bridges across networks. In cross-platform user identification tasks, these anchors typically correspond to accounts known to belong to the same user across different social media platforms. Existing methods propagate the correspondence information carried by the anchors hop-by-hop across the network, spreading alignment constraints from the anchors to the remaining nodes in the network, and then determining the correspondence based on the node similarity obtained after propagation. Furthermore, when extracting node structural information, existing methods typically extract the structural features of nodes independently within each network and learn a low-dimensional representation vector for each node accordingly, then establish correspondences by comparing the similarity between node representations across networks.

[0004] Existing research on unsupervised network alignment has the following three shortcomings:

[0005] On the one hand, the discriminativeness of node representations is insufficient: some existing methods adopt symmetrical message passing when learning node representations, treating nodes in the neighborhood almost equally, failing to reflect the differences in structural importance of different nodes. The learned node representations have limited discriminativeness, which is the basis for subsequent clustering and pseudo-anchor selection. Insufficient discriminativeness will directly affect the accuracy of node matching, making it easier to confuse accounts with similar structures but belonging to different users in the cross-platform user identification process, thereby reducing the reliability of network alignment.

[0006] On the other hand, it is difficult to make consistent and reliable comparisons across network structures: most existing methods extract structural features and learn node representations independently within each network, lacking a unified reference across networks. Node representations are defined only within their own network structure, and cross-network structural similarity can only be inferred indirectly. After introducing high-confidence corresponding node anchors, the method mainly relies on hop-by-hop propagation on the network to spread the corresponding constraints. However, these constraints decay rapidly with topological distance, making it difficult for nodes far from the anchors to obtain effective guidance. Expanding the propagation range will also exacerbate information decay (i.e., oversmoothing), failing to provide a reliable alignment basis for distant nodes. This makes it difficult to establish reliable account correspondences for users located at the network edge or with few friend relationships during cross-platform user identification, easily leading to mismatches or missed matches.

[0007] Furthermore, there are shortcomings in the screening of high-confidence correspondences (pseudo-anchors) in unsupervised network alignment. In unsupervised scenarios, there is a lack of labeled correspondences, requiring automatic screening of pseudo-anchors from a huge candidate space composed of all source and target nodes. However, the candidate space is large and easily contaminated by structurally unrelated noisy node pairs. Existing methods lack effective constraint mechanisms to narrow the search range and eliminate noise interference, making it difficult to guarantee the reliability and screening efficiency of the selected pseudo-anchors. This makes it difficult for a large number of candidate accounts to obtain reliable matching constraints during cross-platform user identification, thereby reducing the overall network alignment performance.

[0008] In summary, existing unsupervised network alignment methods suffer from problems such as insufficient discriminative power of node representations, lack of unified references across network structures, and low reliability of pseudo-anchor point screening. As a result, nodes far from anchor points are difficult to obtain accurate and stable alignment criteria, making it difficult to meet the application requirements of cross-platform user recognition tasks for network alignment accuracy. Summary of the Invention

[0009] The purpose of this invention is to provide an unsupervised network alignment method and device for cross-platform user identification. By using a set of high-confidence pseudo-anchor pairs as a shared structural reference across networks, each node is encoded with its topological distance to all pseudo-anchors in the network. This allows nodes from different networks to be mapped to the same reference and directly compared. Since the node encoding is directly calculated from the network topology without hop-by-hop message propagation, the structural coordinates of nodes far from the pseudo-anchors change more smoothly with topological distance. The method comprises three coupled parts: the first part learns discriminative node representations through centrality-aware asymmetric message passing and, after clustering in the representation space, selects high-confidence pseudo-anchors within corresponding clusters; the second part uses the pseudo-anchors as shared references for coordinate encoding, calculates structural similarity, and fuses it with embedding similarity to obtain the final alignment result; the third part uses the alignment result to guide the clustering optimization of the target network, making the pseudo-anchors selected in the next round more reliable. These three parts iterate repeatedly until a preset number of iterations is reached, outputting the final node correspondence.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] An unsupervised network alignment method for cross-platform user identification includes the following steps:

[0012] Obtain two networks to be aligned, one of which is the source network and the other is the target network;

[0013] Extract the node set, edge set, and initial node features of the two networks to be aligned, and preprocess the initial node features;

[0014] The preprocessed initial node features are input into the graph neural network encoder to obtain the node representations encoded by the source network and the target network.

[0015] The encoded node representations of the source network and the target network are subjected to k-means clustering to obtain the cluster division of the source network and the target network. Then, the mean of the node representations within the source network cluster and the target network cluster is used as the cluster center. The similarity between the center of the source network cluster and the target network cluster is calculated based on cosine similarity. For each source network cluster, the target network cluster with the highest similarity is selected as the corresponding cluster. Within each pair of corresponding clusters, the embedding similarity matrix of source network nodes and target network nodes is calculated. Node pairs with scores higher than the threshold are selected as the pseudo-anchor point set based on node similarity.

[0016] For any node in the source network and the target network, calculate the shortest path distance to all pseudo-anchor points in the network, and obtain continuous structural coordinates by mapping through the distance decay function. Perform L2 normalization on the structural coordinates, and use the inner product of the normalized structural coordinates of the source network and the target network as the global structural similarity.

[0017] For any node in the source network and the target network, the mean of its one-hop neighborhood representation is calculated as the local structural context, and the local structural offset is obtained based on the difference between the node representation and the mean. The local structural offset is L2 normalized, and the inner product of the normalized local structural offsets of the source node and the target node is used as the local structural similarity.

[0018] The global structural similarity and local structural similarity are smoothly fused to obtain the structural similarity matrix. The structural similarity matrix is ​​then fused with the embedding similarity to obtain the final alignment matrix. The embedding similarity is calculated by the embedding similarity of the node representation matrix of the source network and the target network.

[0019] Based on the final alignment matrix, the target network node with the highest similarity to each source network node is selected as the corresponding node.

[0020] Preferably, the preprocessing of the initial node features specifically includes the following steps: mapping the initial node features to a fixed-dimensional representation space and then performing L2 normalization.

[0021] Preferably, the method further includes the following steps: calculating multiple centrality indices for each node based on the topology, and obtaining the comprehensive centrality Cen(v) by averaging the multiple centrality indices, wherein the centrality indices include degree centrality, proximity centrality, betweenness centrality, and eigenvector centrality;

[0022] For each directed edge from node i to node j, the graph neural network encoder obtains the asymmetric propagation weights based on the difference in the combined centrality Cen(v) between nodes i and j.

[0023]

[0024] Then, within the neighborhood of each target node, the asymmetric propagation weights are normalized. The normalized asymmetric propagation weights are then aggregated with the neighbor representations of the node to obtain the aggregated message of node j:

[0025]

[0026] in (j) is the set of neighbors of node j. The normalized asymmetric propagation weights, The representation of neighbor node i is then updated through a shared nonlinear transformation:

[0027] .

[0028] Preferably, the various centrality indices of each node are normalized to [0,1], and the asymmetric propagation weights are clipped to the interval [1-σ,1+σ], where σ>0.

[0029] Preferably, for each pseudo-anchor node, the asymmetric propagation weight to all neighboring nodes j is increased by a constant. :

[0030]

[0031] After the pseudo-anchor weighted propagation is completed, update the node embedding representations of the source network and the target network. , Based on the updated node representations, the embedding similarity matrix is ​​calculated:

[0032] .

[0033] Preferably, the source network structure coordinates are:

[0034]

[0035] in, Let be the node of the nth pair of pseudo-anchors in the source network, and u be the node in the source network;

[0036] The local structural offset of the source network is:

[0037]

[0038] in, Let x be the node representation of node u, and let x be the neighboring nodes of node u. Let x be the node representation of node x.

[0039] Preferably, the structural similarity matrix is:

[0040]

[0041] in, For global structural similarity across network node pairs (u,v), For the local structural similarity of cross-network node pairs (u,v), φ and κ are hyperparameters that control the relative contributions of the global and local networks.

[0042] Preferably, the method further includes the following step: based on the source network cluster In the context of node u, its target node is v, and the corresponding cluster... In the target network, the cluster assignment of node v is expressed in probabilistic form as follows: And it is updated through the following consistency constraints:

[0043]

[0044] Where c(u) is the cluster index to which the source node u belongs, and M is the final alignment matrix. The set corresponding to the Top-1;

[0045] Under the new corresponding cluster constraints, the pseudo-anchor user set is re-selected, and the structural coordinate construction, local offset calculation and similarity matrix fusion steps are re-executed based on the updated pseudo-anchor set. The node representation, structural similarity matrix and target cluster partitioning results are jointly updated to ensure that the cluster affiliation of the matched user in the source network is consistent with its corresponding cluster in the target network.

[0046] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the above-described unsupervised network alignment method for cross-platform user identification.

[0047] In another aspect, the present invention provides an unsupervised network alignment device for cross-platform user identification, comprising:

[0048] Memory, used to store software applications.

[0049] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the unsupervised network alignment method for cross-platform user identification described above.

[0050] To address the issue of insufficient discriminative power in existing node representations, this invention proposes an asymmetric message passing mechanism based on centrality awareness. By applying asymmetric weighting to the neighborhood based on node centrality differences, core and active users in the social network can play a greater role in message propagation, thereby learning more discriminative node representations.

[0051] To address the technical problem of the lack of a unified reference across networks, which makes it difficult for nodes far from pseudo-anchors to obtain effective alignment guidance, this invention proposes a coordinate-based structural alignment mechanism. Using high-confidence pseudo-anchors as shared references, each node is encoded as its topological distance to each pseudo-anchor, enabling direct comparison between nodes in different networks under a unified reference. For user nodes with fewer social relationships or located at the network edge, even if they are far from pseudo-anchors, a reliable correspondence can be established based on their relative positions with multiple shared reference points, thus providing a stable alignment basis for a wider range of nodes.

[0052] To address the issues of a large candidate space and susceptibility to noise interference in pseudo-anchor selection in unsupervised scenarios, this invention proposes a cluster-constrained pseudo-anchor selection mechanism. First, a cross-network cluster-level correspondence is established, and pseudo-anchor selection is performed only within corresponding clusters. This effectively narrows the search range, reduces interference from irrelevant node pairs, and improves the reliability of pseudo-anchor selection, thereby enhancing the accuracy of cross-platform user identification tasks. Detailed Implementation

[0053] The present invention will be further described below with reference to specific embodiments.

[0054] This invention discloses an unsupervised network alignment method for cross-platform user identification, comprising the following steps:

[0055] S1. Obtain two networks to be aligned, one of which is the source network and the other is the target network;

[0056] Specifically, this embodiment takes cross-network user alignment on a social platform as an example, using the publicly available DoubanOnline-Offline dataset to illustrate the specific implementation process of the present invention. In this social platform, the same group of users have both online interactive relationships (such as following and commenting) and offline activity relationships (such as participating in local events together), respectively constituting an online network G. s With offline network G t Among them, online network G s For the source network, the offline network G t The target network is defined as follows: Nodes in both networks are users, but their connection relationships differ. The objective of this invention is to identify the same user in both networks without any known account mappings.

[0057] S2. Extract the node set, edge set, and initial node features of the two networks to be aligned, and preprocess the initial node features;

[0058] Specifically, firstly, two networks to be aligned are given. and Extract its node set and edge set, and construct the adjacency matrix respectively. To represent the network structure, the initial features of the nodes are mapped to a fixed-dimensional representation space and then subjected to L2 normalization, denoted as . This is to eliminate scale differences between different networks.

[0059] (1) Centrality-aware message passing and intra-cluster pseudo-anchor screening strategy

[0060] This part is used to learn discriminative node representations and, based on this, to select high-confidence pseudo-anchor points from the candidate space, providing a reliable structural reference for subsequent coordinate alignment.

[0061] In cross-platform user identification, core users typically possess more stable social relationships and richer structural information; therefore, their neighborhood information can provide more reliable structural references for other user nodes. This invention introduces node centrality as a structural prior and applies asymmetric weighting to neighborhood messages.

[0062] S3. Calculate multiple centrality indices for each node based on the topology (all normalized to [0,1]), and obtain the comprehensive centrality Cen(v) by averaging the multiple centrality indices. The centrality indices are pre-calculated by the network topology without introducing learnable parameters. The centrality indices include degree centrality, proximity centrality, betweenness centrality, and eigenvector.

[0063] Specifically, the eigenvector centrality is: Ac = λc, where A is the adjacency matrix, λ is the largest eigenvalue of the adjacency matrix, and c is the eigenvector centrality vector, where c is the eigenvector centrality vector. i Indicate the eigenvector centrality of node i;

[0064] Overall centrality reflects the structural importance of users in the network. Core users with high attention and activity have higher centrality and are used for subsequent centrality-weighted message passing. Ultimately, this forms the source and target network inputs with consistent structure and features, providing basic data support for subsequent clustering, pseudo-anchor selection, and cross-network alignment.

[0065] S4. Input the preprocessed initial features of the nodes into the graph neural network encoder to obtain the node representations encoded by the source network and the target network.

[0066] Specifically, for each directed edge from node i to node j, the graph neural network encoder obtains the asymmetric propagation weights based on the difference in the combined centrality Cen(v) between nodes i and j:

[0067]

[0068] Since Cen(v) takes values ​​in the range [0,1], all propagation weights are always positive, and messages from higher-centrality nodes are amplified when propagating to lower-centrality nodes, while backpropagation is suppressed. To avoid excessive amplification by high-centrality nodes and disruption of local structural consistency, the weights are pruned to the range [0,1]. , where σ>0 controls the maximum adjustment range and is normalized in the neighborhood of each target node to ensure numerical stability;

[0069] Then, within the neighborhood of each target node, the asymmetric propagation weights are normalized. The normalized asymmetric propagation weights are then aggregated with the neighbor representations of the node to obtain the aggregated message of node j:

[0070]

[0071] in (j) is the set of neighbors of node j. The normalized asymmetric propagation weights, The representation of neighbor node i is then updated through a shared nonlinear transformation:

[0072] .

[0073] S5, Node representation after encoding of source network and target network , Perform k-means clustering separately to obtain the cluster partitions of the source and target networks. , Each cluster consists of nodes with similar structural features. The mean value of the node representations in the source network cluster and the target network cluster is used as the cluster center. The similarity between the center of the source network cluster and the target network cluster is calculated based on cosine similarity. The target network cluster with the highest similarity is selected as the corresponding cluster for each source network cluster. Within each pair of corresponding clusters, the embedding similarity matrix of source network nodes and target network nodes is calculated. Node pairs with scores higher than the threshold are selected as the pseudo-anchor point set based on node similarity.

[0074] Specifically, after obtaining the node representations of the source network and the target network, this invention first establishes cross-network correspondences at the cluster level, and then restricts the selection of pseudo-anchor points to corresponding clusters. That is, k-means clustering is performed in the representation spaces of the source network and the target network respectively to obtain the representations of the source network. Clusters and target networks There are several clusters, with the cluster number chosen as the square root of the number of nodes, and the cluster center being the mean value represented by the nodes within the cluster.

[0075]

[0076]

[0077] Then, the similarity matrix between the source network cluster and the target network cluster center is calculated based on cosine similarity. This is used to measure the structural correspondence between network clusters, and based on this, for each source cluster... Select the target cluster with the highest similarity as the corresponding cluster. This establishes a one-to-one correspondence at the cluster level. Subsequently, pseudo-anchor points are filtered only within corresponding clusters, restricting target candidates to matching target clusters. This ensures that the selected node pairs are both similar at the node level and located in corresponding clusters across the network, thereby narrowing the search scope from all node pairs to corresponding clusters, reducing interference from noisy node pairs, and improving the structural consistency and stability of cross-network matching. Finally, the cluster partitioning results corresponding across the network are obtained.

[0078] After establishing the cluster-level correspondence, a set of pseudo-anchor users is generated under the corresponding cluster constraints, and the node representations are weighted and aligned based on the pseudo-anchor users.

[0079] Specifically, in each pair of corresponding clusters Within the set of pseudo-anchor points, node pairs with higher similarity scores are selected. To more fully utilize the pseudo-anchor points to guide feature learning, this invention further applies propagation enhancement to the pseudo-anchor nodes. In the next round of message passing, each pseudo-anchor user node is given additional weight during graph propagation. :

[0080]

[0081] This enhances its impact on the representation updates of neighboring nodes, thereby strengthening the propagation of consistent information across the network;

[0082] After the pseudo-anchor weighted propagation is completed, the node embedding representations of the source network and the target network are updated. Based on the updated node representations, the embedding similarity matrix is ​​calculated.

[0083]

[0084] (2) Pseudo-anchor guided coordinate-based network alignment

[0085] Embedding similarity alone is often insufficient to determine node correspondence, especially for nodes far from high-confidence anchors. Their embedding is limited by the receptive field of their local neighborhood, making it difficult to maintain a globally consistent structural position. Existing methods often rely on hop-by-hop propagation of anchor constraints to introduce structural information, but this constraint decays with distance, weakening the alignment guidance for distant nodes. To address this, this invention uses high-confidence pseudo-anchors as a shared structural coordinate system across the network, representing each node as its topological distance to each pseudo-anchor, thus providing a consistent structural reference. Furthermore, by leveraging the inherent global topological relationship between the two networks, the alignment guidance is extended beyond the local neighborhood of the anchors. For ordinary user nodes located at the network edge or with fewer social relationships, even without direct correspondence information, a stable structural representation can be obtained based on their relative topological position with multiple shared reference points.

[0086] S6. For any node in the source network and the target network, calculate the shortest path distance to all pseudo-anchor points in the network, and obtain continuous structural coordinates by mapping through the distance decay function. Perform L2 normalization on the structural coordinates, and use the inner product of the normalized structural coordinates of the source network and the target network as the global structural similarity.

[0087] Specifically, let the set of pseudo-anchor pairs be... Where m is the number of pseudo-anchor pairs, and Let be the nodes of the nth pair of pseudo-anchor points in the source and target networks, respectively. Since these node pairs have high cross-network correspondence confidence, they can be used to define a shared structural coordinate system between the two networks. The positions of all other nodes are measured against this system. For any node in the source network... Calculate the shortest path distance from it to all pseudo-anchor points in this network, and map it to continuous structural coordinates using a distance decay function:

[0088]

[0089] Where d(·,·) represents the shortest path distance.

[0090] The structural coordinates of the target network nodes are constructed similarly. These coordinates directly encode the relative topological distance from the node to the shared set of pseudo-anchor points, making the node positions in different networks naturally comparable. Subsequently, L2 normalization is applied to the structural coordinates to eliminate scale differences and ensure numerical stability. The inner product of the normalized coordinates is used as the global structural similarity, reflecting whether two nodes occupy consistent positions in the global topology. The global structural similarity for any cross-network node pair (u,v) is:

[0091]

[0092] in, For the normalized source network structure coordinates, These are the normalized coordinates of the target network structure.

[0093] S7. For any node in the source network and the target network, calculate the mean of its one-hop neighborhood representation as the local structural context, and obtain the local structural offset based on the difference between the node representation and the mean. Perform L2 normalization on the local structural offset, and use the inner product of the normalized local structural offsets of the source node and the target node as the local structural similarity.

[0094] Specifically, global coordinates characterize the macroscopic topological location of a node, but do not reflect its local connectivity within its neighborhood. This makes it difficult to distinguish nodes that are macroscopically close but have different local structures. To address this, this invention uses local structure offset as a supplement. For a source network node u, the mean of its one-hop neighborhood representation is taken as the local structure context, and the local structure offset is defined as the difference between the node representation and this mean.

[0095] For any user node Based on its updated node embedding representation The local structural offset vector of a node is defined as follows:

[0096]

[0097] in, Let x be the node representation of node u, and let x be the neighboring nodes of node u. Let x be the node representation; this offset removes the translation components shared by the neighborhood, highlights the functional differences in the local structure, and characterizes the structural role of the node relative to its neighborhood.

[0098] Similarly, the local offset vector of node v in the target network is calculated by performing L2 normalization on the local structural offsets, and using the inner product of the normalized local offsets as the local structural similarity:

[0099]

[0100] in, For the normalized local offset of the source network, The target network local offset is normalized.

[0101] S8. Smoothly fuse the global structural similarity and the local structural similarity to obtain the structural similarity matrix. Then fuse the structural similarity matrix with the embedding similarity to obtain the final alignment matrix. The embedding similarity is calculated by the embedding similarity of the node representation matrix of the source network and the target network.

[0102] Specifically, the structural similarity matrix:

[0103]

[0104] in , To control the hyperparameters of global and local relative contributions, take... > The formula makes the fusion biased towards global structural information a smooth maximum operator. This formula retains complementary information while highlighting the dominant structural information, so that local similarity can play a corrective role on the correspondence dominated by global structure.

[0105] The resulting structural similarity matrix is ​​then fused with the embedding similarity matrix to obtain the final alignment matrix:

[0106]

[0107] Where λ is used to balance the contributions of embedding similarity and structural similarity in alignment;

[0108] This matrix integrates node attribute semantic information and multi-scale structural information as the final cross-network matching basis, and selects the target node with the highest similarity to each source node as the corresponding node.

[0109] (3) Clustering optimization guided by alignment results

[0110] The aforementioned coordinate-based alignment depends on the quality of the pseudo-anchor points, which are selected from the cluster-level correspondences obtained through clustering. To ensure that node-level alignment and cluster-level structure mutually promote each other, this invention, while keeping the source network clustering fixed, guides the optimization of the target network clustering distribution with the alignment results, gradually aligning it with the cluster structure of the source network, thereby selecting more reliable pseudo-anchor points in subsequent iterations.

[0111] S9, Based on the source network cluster In the context of node u, its target node is v, and the corresponding cluster... In the target network, the cluster assignment of node v is expressed in probabilistic form as follows: And it is updated through the following consistency constraints:

[0112]

[0113] Where c(u) is the cluster index to which the source node u belongs, and M is the final alignment matrix. The set corresponding to the Top-1 is where Top-1 is the point with the highest score in each row of the similarity matrix, that is, the target node most similar to each source node. This loss encourages each corresponding target node to assign a higher probability to its corresponding target cluster. During optimization, the clustering distribution of the source network remains fixed, and the clustering distribution of the target network is updated by gradient descent. After each update, the probability distribution of the target node is normalized.

[0114] Under the new corresponding cluster constraints, the pseudo-anchor user set is re-selected, and the structural coordinate construction, local offset calculation and similarity matrix fusion steps are re-executed based on the updated pseudo-anchor set. The node representation, structural similarity matrix and target cluster partitioning results are jointly updated so that the cluster affiliation of the matching user in the source network is consistent with its corresponding cluster in the target network. The above process is executed cyclically according to the preset iteration strategy until the maximum number of iterations is reached.

[0115] Through the aforementioned iterative optimization, the cluster structure of the target network gradually incorporates cross-network correspondence information, reinforcing node-level alignment and cluster-level structure. The refined target network clustering is used for the next round of intra-cluster pseudo-anchor selection, resulting in more reliable pseudo-anchor points. These more reliable pseudo-anchor points further improve coordinate encoding and alignment quality. This process repeats, with the quality of pseudo-anchor points and alignment quality gradually improving with each iteration until the preset number of iterations (it) is reached, at which point the final node correspondence is output.

[0116] S10. Based on the final alignment matrix, select the target network node with the highest similarity to each source network node as the corresponding node, and output the corresponding user pair in the online network and the offline network. The obtained correspondence can be used to integrate the user's online interests and offline behavior to build a more complete user profile.

[0117] To illustrate the effectiveness of the invention, experiments were conducted on four publicly available network alignment benchmark datasets, which cover four different types of networks: social networks, academic collaboration networks, film networks, and biomolecular interaction networks.

[0118] Douban Online-Offline: This dataset is a commonly used benchmark dataset for social network alignment, containing online interaction networks and offline friend networks collected from the Douban platform. Nodes represent users, and edges represent social interactions or friend relationships between users within each network; some users appear in both networks simultaneously, and the actual correspondence between these shared users serves as the basis for network alignment.

[0119] ACM-DBLP: This dataset is a cross-network dataset constructed from two academic collaboration networks, one sourced from the ACM Digital Library and the other from the DBLP document repository. Nodes represent authors, and edges represent collaboration relationships. Since many researchers have published in both repositories, a set of true author correspondences can be obtained, which is used to evaluate cross-network alignment performance.

[0120] Allmv-Tmdb: This dataset is a cross-domain film network dataset constructed by two film information platforms. Nodes in both networks represent films, and edges indicate that two films share at least one actor, reflecting the collaborative relationships between actors in different films. Films appearing in both databases provide a true correspondence, used to evaluate the accuracy of alignment algorithms.

[0121] PPI: This dataset is a protein-protein interaction network dataset derived from biological databases. Nodes represent proteins, and edges represent functional interactions between proteins. The source and target networks contain the same protein nodes, and known protein correspondences serve as a reliable basis for evaluating alignment performance.

[0122] In terms of experimental setup, this invention is implemented based on PyTorch 1.12.1 and PyTorch Geometric, employing a 3-layer graph neural network with a node representation dimension of 1024, and the number of clusters is the square root of the total number of nodes. Key hyperparameters such as the number of iterations, structural similarity weight λ, and centrality adjustment magnitude δ are determined through sensitivity analysis. All real node correspondences are used only for evaluation and not for training to ensure an unsupervised setup. Experiments were conducted on a server equipped with four NVIDIA GeForce RTX 3090 GPUs (each with 24 GB of VRAM). This invention was compared with six representative network alignment methods, using the mean average precision (MAP) and Precision@k (k values ​​are 1, 5, and 10), commonly used in network alignment, as evaluation metrics; higher values ​​indicate more accurate alignment.

[0123] To verify the effectiveness of this invention, it was compared with the nine methods mentioned above on four benchmark datasets, and the results are shown in Table 1. The symbol "-" indicates that the method cannot give effective results under the current experimental settings, and the superscript 1 indicates that the method uses 10% of the true correspondence as prior knowledge.

[0124] Table 1 Comparison of Experimental Results

[0125]

[0126] As shown in the table, this invention achieves optimal performance across all metrics on the four datasets. On the DoubanOnline-Offline dataset, the MAP of this invention reaches 0.6771, significantly higher than SLOTAlign's 0.6129. Precision@1, Precision@5, and Precision@10 also reach 0.5796, 0.8014, and 0.8685 respectively, all of which are optimal. On the ACM-DBLP dataset, this invention improves the MAP from SLOTAlign's 0.7376 to 0.7626, and maintains a leading position across all Precision@k metrics. On the Allmv-Tmdb dataset, the MAP and Precision@1 of this invention reach 0.9284 and 0.9090 respectively, outperforming the strong baseline SLOTAlign (MAP of 0.9161) and significantly exceeding KNN, which relies solely on embedding similarity (MAP of 0.4150 and Precision@1 of 0.3239). On the PPI dataset, the MAP of this invention reaches 0.9107, which is also superior to the comparison methods.

[0127] The above results validate the effectiveness of this invention in cross-network node alignment. This is mainly due to two aspects: First, the centrality-aware asymmetric message passing adjusts message propagation based on the global importance of nodes, enabling the model to distinguish nodes with similar local structures but different global roles, thus learning more discriminative representations. Second, the alignment in this invention not only relies on the learned embeddings but also utilizes the topological distance from each node to the high-confidence pseudo-anchor point as a shared structural coordinate across the network, providing an effective supplement when embedding similarity alone is insufficient to fully reflect cross-network structural consistency. The combined effect of these two aspects enables this invention to achieve consistent alignment performance improvements across different types of networks.

[0128] To verify the contribution of the three core modules of this invention to alignment accuracy, ablation tests were conducted using the Douban Online-Offline and ACM-DBLP datasets. The hardware and software environment, evaluation metrics, and comparative experiments were kept consistent. Four control groups were set up: complete scheme, removal of the centrality module, removal of the clustering module, and removal of the structural similarity module. The experimental results are shown in Table 2.

[0129] Table 2 Ablation Experiment Results

[0130]

[0131] Data shows that centrality-aware message passing, clustering pseudo-anchor construction, and reference-based structural similarity are all key components for improving alignment accuracy. Optimal results are achieved through their synergy. The structural similarity module shows the most significant gain, compensating for insufficient embedding representations by using polynomial-decreasing topological gain. The centrality mechanism differentiates and passes features, improving node embedding discrimination. Clustering filters high-confidence pseudo-anchors, ensuring the stability of the structural coordinate baseline. In the noisier Douban network, the performance decline is more pronounced after removing modules. While the performance decline is smaller in the structurally regular ACM-DBLP network, each module consistently shows positive performance gains, adapting to different network topologies.

[0132] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the above-described unsupervised network alignment method for cross-platform user identification.

[0133] In another aspect, the present invention provides an unsupervised network alignment device for cross-platform user identification, comprising:

[0134] Memory, used to store software applications.

[0135] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the unsupervised network alignment method for cross-platform user identification described above.

Claims

1. An unsupervised network alignment method for cross-platform user identification, characterized in that, Includes the following steps: Obtain two networks to be aligned, one of which is the source network and the other is the target network; Extract the node set, edge set, and initial node features of the two networks to be aligned, and preprocess the initial node features; The preprocessed initial node features are input into the graph neural network encoder to obtain the node representations encoded by the source network and the target network. The encoded node representations of the source network and the target network are subjected to k-means clustering to obtain the cluster division of the source network and the target network. Then, the mean of the node representations within the source network cluster and the target network cluster is used as the cluster center. The similarity between the center of the source network cluster and the target network cluster is calculated based on cosine similarity. For each source network cluster, the target network cluster with the highest similarity is selected as the corresponding cluster. Within each pair of corresponding clusters, the embedding similarity matrix of source network nodes and target network nodes is calculated. Node pairs with scores higher than the threshold are selected as the pseudo-anchor point set based on node similarity. For any node in the source network and the target network, calculate the shortest path distance to all pseudo-anchor points in the network, and obtain continuous structural coordinates by mapping through the distance decay function. Perform L2 normalization on the structural coordinates, and use the inner product of the normalized structural coordinates of the source network and the target network as the global structural similarity. For any node in the source network and the target network, the mean of its one-hop neighborhood representation is calculated as the local structural context, and the local structural offset is obtained based on the difference between the node representation and the mean. The local structural offset is L2 normalized, and the inner product of the normalized local structural offsets of the source node and the target node is used as the local structural similarity. The global structural similarity and local structural similarity are smoothly fused to obtain the structural similarity matrix. The structural similarity matrix is ​​then fused with the embedding similarity to obtain the final alignment matrix. The embedding similarity is calculated by the embedding similarity of the node representation matrix of the source network and the target network. Based on the final alignment matrix, the target network node with the highest similarity to each source network node is selected as the corresponding node.

2. The unsupervised network alignment method for cross-platform user identification according to claim 1, characterized in that, Preprocessing the initial features of nodes specifically includes the following steps: mapping the initial features of nodes to a fixed-dimensional representation space and then performing L2 normalization.

3. The unsupervised network alignment method for cross-platform user identification according to claim 1, characterized in that, It also includes the following steps: Based on the topology, multiple centrality indices are calculated for each node, and the comprehensive centrality Cen(v) is obtained by averaging the multiple centrality indices. The centrality indices include degree centrality, proximity centrality, betweenness centrality, and eigenvector centrality. For each directed edge from node i to node j, the graph neural network encoder obtains the asymmetric propagation weights based on the difference in the combined centrality Cen(v) between nodes i and j. Then, within the neighborhood of each target node, the asymmetric propagation weights are normalized. The normalized asymmetric propagation weights are then aggregated with the neighbor representations of the node to obtain the aggregated message of node j: in (j) is the set of neighbors of node j. The normalized asymmetric propagation weights, The representation of neighbor node i is then updated through a shared nonlinear transformation: 。 4. The unsupervised network alignment method for cross-platform user identification according to claim 3, characterized in that, The centrality indices of each node are normalized to [0,1], and the asymmetric propagation weights are clipped to the interval [1-σ,1+σ], where σ>0.

5. The unsupervised network alignment method for cross-platform user identification according to claim 3, characterized in that, For each pseudo-anchor node, increase its asymmetric propagation weight to all neighboring nodes j by a constant. : After the pseudo-anchor weighted propagation is completed, update the node embedding representations of the source network and the target network. , Based on the updated node representations, the embedding similarity matrix is ​​calculated: 。 6. The unsupervised network alignment method for cross-platform user identification according to claim 1, characterized in that, The source network structure coordinates are: in, Let be the node of the nth pair of pseudo-anchors in the source network, and u be the node in the source network; The local structural offset of the source network is: in, Let x be the node representation of node u, and let x be the neighboring nodes of node u. Let x be the node representation of node x.

7. The unsupervised network alignment method for cross-platform user identification according to claim 1, characterized in that, The structural similarity matrix is: in, For global structural similarity across network node pairs (u,v), For the local structural similarity of cross-network node pairs (u,v), φ and κ are hyperparameters that control the relative contributions of the global and local networks.

8. The unsupervised network alignment method for cross-platform user identification according to claim 1, characterized in that, It also includes the following steps: According to the source network cluster In the context of node u, its target node is v, and the corresponding cluster... In the target network, the cluster assignment of node v is expressed in probabilistic form as follows: And it is updated through the following consistency constraints: Where c(u) is the cluster index to which the source node u belongs, and M is the final alignment matrix. The set corresponding to the Top-1 Under the new corresponding cluster constraints, the pseudo-anchor user set is re-selected, and the structural coordinate construction, local offset calculation and similarity matrix fusion steps are re-executed based on the updated pseudo-anchor set. The node representation, structural similarity matrix and target cluster partitioning results are jointly updated to ensure that the cluster affiliation of the matched user in the source network is consistent with its corresponding cluster in the target network.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the steps of the unsupervised network alignment method for cross-platform user identification as described in any one of claims 1 to 8.

10. An unsupervised network alignment device for cross-platform user identification, characterized in that, include: Memory, used to store software applications. A processor for executing the software application, wherein each program of the software application correspondingly performs the steps in the unsupervised network alignment method for cross-platform user identification as described in any one of claims 1 to 8.

Images