Cross-Social Network User Identity Recognition Method and System Based on Spatial Consistency Representation

Abstract

The present invention discloses a method and system for cross-social network user identity recognition based on spatial consistency representation. The method includes: applying a graph convolutional neural network to encode the source network and the target network from the structural space into a unified vector representation space; using an inner product decoder to reconstruct the network structures of the source network and the target network, and constructing a reconstruction loss by minimizing the difference between the reconstructed network and the original network; constructing positive samples and negative samples based on known anchor links, and constructing a contrast loss by the constraint that the distance between positive samples is closer than that between negative samples in the vector representation space; optimizing the spatial consistency representation of nodes by the joint constraint of the reconstruction loss and the contrast loss, so that the learned node feature representation is more conducive to solving the problem of cross-social network user identity recognition. The present invention encodes different social networks into a unified vector representation space, makes the user node representations in different networks have spatial consistency, and improves the accuracy of cross-social network user identity recognition.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for identifying user identities across social networks based on spatial consistency representation. Background Art

[0002] Cross-social network user identity recognition aims to discover the corresponding relationship between the accounts of the same user in different social networks, which is of great significance for cross-domain information dissemination and cross-platform friend recommendations.

[0003] Once the problem of cross-social network user identity recognition was proposed, it has attracted widespread attention from scholars in academia and industry, and many excellent research results have emerged, which can be mainly divided into three categories. The first category is methods based on user attribute features such as username, avatar, and geographic location. The attribute information of user registration is used to extract features representing user identity. The performance of this type of method depends entirely on the accuracy of attribute information. The second category is methods based on user behavior features such as tweet content, writing habits, and activity trajectories. User identity features are obtained from the behavioral information generated when users participate in social activities. This type of method is mainly limited by the richness of user behavior information. The third category is methods based on user structural features. Most of these methods are based on the structural consistency assumption, that is, the same user often has a consistent relationship structure in different social networks, and local structural features and global structural features are extracted from social relationships to characterize user identity. From the perspective of development, structure-based methods can be roughly divided into two categories: matrix decomposition-based methods, which use matrix decomposition from the perspective of traditional graphs to transform the user identity recognition problem into a node matching problem between two or more networks; and graph representation-based methods, which use graph embedding and graph neural networks to extract the structural feature representation of users and transform the user identity recognition problem into a similarity measurement problem between users. In recent years, many scholars have also tried to combine user attribute information or user behavior information with network structure information, and carried out related research based on multi-dimensional information union. However, how to accurately extract social network structural features is still the most critical step in this type of method, so methods based on user structural features are still the mainstream of current research. In the present invention, the present invention conducts relevant research in a scenario where only social network relationship structures are available.

[0004] The emergence of graph representation learning has brought a new research boom to cross-social network user identity recognition. The graph representation-based method consists of two main parts: First, apply graph representation learning to preserve the topological structure information of the network and learn the low-dimensional dense vector representation of nodes; then identify the corresponding relationships between nodes across social networks. Man et al. [Man T, Shen H, Liu S, et al. Predict anchor links across social networks via an embedding approach[C] / / Ijcai. 2016, 16: 1823-1829.] proposed a two-stage PALE model. First, use network representation learning to obtain the identity feature representation of users, and then use the known anchor links as supervision information to learn a mapping function for predicting unknown anchor links. Different from PALE, the GraphUIL model [Zhang W, Shu K, Liu H, et al. Graph neural networks for user identity linkage[J]. arXiv preprint arXiv:1903.02174, 2019.] optimizes learning user feature representation and user identity recognition in a unified framework. First, apply graph neural networks to capture local and global structure information of social networks, map the source network and the target network to different vector spaces respectively, and then use a multi-layer perceptron model to map one vector space to another vector space. The introduction of graph representation technology enables such methods to efficiently process large-scale networks. Since these methods represent the nodes of the source network and the target network independently, the node representations in different networks do not have spatial consistency, that is, the node representations of the source network and the target network are not in the same vector representation space and cannot be directly compared. Therefore, a mapping function is needed to transform the two independent vector representation spaces into a consistent vector space so that the node representations from different networks satisfy comparability. However, the introduction of the mapping function will bring additional errors, and the mapping function learned only using limited known anchor link information may not be applicable to other nodes. Therefore, it is very necessary to learn node representations with spatial consistency. Hong et al. proposed the DANA [Hong H, Li X, Pan Y, et al. Domain-adversarial network alignment[J]. IEEE Transactions on Knowledge and Data Engineering, 2020: 3211-3224.] model, which incorporates the domain adversarial principle in the process of using graph convolutional neural networks to learn node representations, aiming to filter out domain-related features and attempt to obtain consistent representations of nodes in different networks.In terms of the accuracy of user identity recognition, DANA has achieved good results. However, the node representations learned by the DANA method cannot well preserve the original structural information of the source network and the target network, and there is still room for improvement. Summary of the Invention

[0005] In view of the problems existing in the existing algorithms based on representation learning, such as inconsistent node representation spaces, inability to well preserve the original structural information of the source network and the target network, and resulting in poor cross-social network user identity recognition effects, the present invention proposes a cross-social network user identity recognition method and system based on spatially consistent representation.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] On the one hand, the present invention proposes a cross-social network user identity recognition method based on spatially consistent representation, including:

[0008] Encoding the source network and the target network from the structural space into a unified vector representation space by stacking multiple graph convolutional layers as an encoder to obtain the spatially consistent representation of user nodes; both the source network and the target network are social networks;

[0009] Using an inner product decoder to reconstruct the network structures of the source network and the target network, and constructing a reconstruction loss by minimizing the difference between the reconstructed network and the original network;

[0010] Taking known anchor links as positive samples, and randomly constructing a certain proportion of negative samples for each positive sample, and constructing a contrast loss by the constraint that the distance of positive samples in the vector representation space is closer than that of negative samples;

[0011] Optimizing the spatially consistent representation of nodes by the joint constraint of the reconstruction loss and the contrast loss;

[0012] Performing cross-social network user identity recognition based on the optimized spatially consistent representation of nodes.

[0013] Further, the expression of the reconstruction loss is:

[0014]

[0015]

[0016] where y s and y t respectively represent the elements in the adjacency matrix A s of the source network and the adjacency matrix A t of the target network, and their values are 0 or 1; and respectively represent the reconstructed adjacency matrix of the source network Reconstruct the adjacency matrix with the target network The elements in; N s And N t respectively represent the number of nodes in the source network and the target network; and U s 、U t respectively represent the node representations output by the encoders of the source network and the target network. (U s ) T 、(U t ) T respectively represent the transposes of U s 、U t . σ is the activation function Sigmoid.

[0017] Furthermore, the contrast loss is:[[]]

[0018]

[0019] where y label represents the label of the sample . The positive sample is 1 and the negative sample is 0; represents the Euclidean distance of the sample ; represents the number of samples in the training sample set; margin is the set threshold.

[0020] Furthermore, the jointly constrained optimization of the spatial consistency representation of nodes using the reconstruction loss and the contrast loss includes:[[]]

[0021] Combining formulas (6), (7), and (8), construct an overall loss function by weighted combination:[[]]

[0022]

[0023] where α is the weight coefficient that controls the ratio of the reconstruction loss and the contrast loss;

[0024] Use the Adam optimizer to minimize the overall loss function (9), and apply the error backpropagation algorithm to achieve iterative optimization to obtain the spatial consistency representation of the nodes based on the optimization.

[0025] On the other hand, the present invention proposes a cross-social network user identity recognition system based on spatial consistency representation, including:[[]]

[0026] A node consistency representation module, which is used to encode the source network and the target network from the structural space into a unified vector representation space by stacking multiple graph convolutional layers as encoders, and obtain the spatial consistency representation of user nodes; both the source network and the target network are social networks;

[0027] The network structure reconstruction module is used to reconstruct the network structures of the source network and the target network by using an inner product decoder, and construct a reconstruction loss by minimizing the difference between the reconstructed network and the original network;

[0028] The contrastive loss construction module is used to take known anchor links as positive samples, and randomly construct a certain proportion of negative samples for each positive sample, and construct a contrastive loss by the constraint that the distance of positive samples is closer than that of negative samples in the vector representation space;

[0029] The joint optimization module is used to optimize the spatial consistency representation of nodes by the joint constraints of the reconstruction loss and the contrastive loss;

[0030] The user identity recognition module is used to perform cross-social network user identity recognition based on the optimized spatial consistency representation of nodes.

[0031] Furthermore, the expression of the reconstruction loss is:

[0032]

[0033]

[0034] where y s and y t represent the elements in the adjacency matrix A s of the source network and the adjacency matrix A t of the target network respectively, and their values are 0 or 1; and represent the elements in the reconstructed adjacency matrix of the source network and the reconstructed adjacency matrix of the target network respectively; N s and N t represent the number of nodes in the source network and the target network respectively; and U s and U t represent the node representations output by the encoders of the source network and the target network respectively, (U s ) T and (U t ) T represent the transposes of U s and U t respectively, and σ is the activation function Sigmoid.

[0035] Furthermore, the contrastive loss is:

[0036]

[0037] where y label represents the label of the sample , the positive sample is 1, and the negative sample is 0; Represents a sample of Euclidean distance; represents the number of samples in the training sample set; margin is the set threshold.

[0038] Furthermore, the joint optimization module is specifically used for:

[0039] Combining formulas (6), (7), and (8), construct an overall loss function in a weighted combination manner:

[0040]

[0041] where α is a weight coefficient for controlling the ratio of the reconstruction loss and the contrast loss;

[0042] Use the Adam optimizer to minimize the overall loss function (9), and apply the error backpropagation algorithm to achieve iterative optimization, obtaining a spatially consistent representation based on the optimized nodes.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] The present invention can encode different social networks into a unified vector representation space, making the user node representations in different networks have spatial consistency. At the same time, the constraint of the reconstruction loss ensures that the learned user node representations are as close as possible to the original network structure features, and the constraint of the contrast loss makes the positive samples closer and closer in the vector space, while the negative samples are farther and farther apart, making the learned user node representations more discriminative, thereby improving the accuracy of cross-social network user identity recognition. Experimental results on three real social network datasets show that the method proposed by the present invention is superior to various existing benchmark algorithms. At the same time, the results of the ablation experiment also illustrate the rationality and effectiveness of the method proposed by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic flowchart of a cross-social network user identity recognition method based on spatially consistent representation according to an embodiment of the present invention;

[0046] Figure 2 is the specific experimental result of precision@k of the embodiment of the present invention on three datasets (Douban-Weibo, Foursquare-Twitter, Dblp17-Dblp19);

[0047] Figure 3 is the MAP result of the embodiment of the present invention on three datasets;

[0048] Figure 4For the P@1 and MAP metrics on the Foursquare-Twitter dataset in the embodiments of the present invention, as the proportion of anchor links changes;

[0049] Figure 5 For the p@1 and MAP on the Foursquare-Twitter dataset in the embodiments of the present invention, as K changes;

[0050] Figure 6 Schematic diagram of the architecture of a cross-social network user identity recognition system based on spatially consistent representation in the embodiments of the present invention. Detailed implementation manners

[0051] The present invention will be further explained and illustrated below in conjunction with the accompanying drawings and specific embodiments:

[0052] 1. Cross-social network user identity recognition framework based on spatially consistent representation

[0053] The present invention uses to represent a social network, where represents the set of N users (nodes) in the social network; represents the set of connection relationships between users in the social network. The present invention uses to represent the adjacency matrix of the social network . If node v i and node v j are connected, then a ij = 1; otherwise a ij = 0. Given two social networks, denoted as the source network and the target network , use to represent the set of known anchor links between the two networks. For the convenience of subsequent description, the present invention gives the definitions of positive samples and negative samples.

[0054] Positive Samples and Negative Samples: The present invention refers to known anchor links as positive samples; for positive samples the present invention refers to the sample and as negative samples. The present invention refers to the distance between the two nodes constituting a sample as the sample distance.

[0055] 1.1 Framework overview

[0056] The present invention models user feature representation learning and user identity recognition in a unified framework, and proposes a cross-social network user identity recognition method CLink based on spatially consistent representation. This method encodes the source network and the target network from the structural space into a unified vector representation space, and jointly optimizes the node representation process using two parts of constraints, that is, it is necessary to ensure that the learned node representations are as close as possible to the original structural features of the source network and the target network, and at the same time, it is also necessary to ensure that the anchor node pairs are closer to each other than the non-anchor node pairs in the unified vector representation space. The overall framework is as shown in Figure 1 and includes four main parts: learning node feature representation, reconstructing the original structure to establish a reconstruction loss, constructing positive and negative samples to establish a contrast loss, and unknown anchor link prediction. The specific steps of this method are as follows:

[0057] Step 1: Node consistency representation based on graph convolutional neural network. As shown in ① of Figure 1 , the present invention stacks multiple graph convolutional layers as an encoder to encode the source network and the target network from the structural space into a unified vector representation space, and obtains the spatial consistency representation of the nodes;

[0058] Step 2: Reconstruct the original network structure and establish a reconstruction loss. After obtaining the node representations in the source network and the target network, the present invention uses an inner product decoder to reconstruct the relationship structure of the original network, and constructs a reconstruction loss based on the difference between the reconstructed network and the original network, as shown in ② of Figure 1 ;

[0059] Step 3: Construct positive and negative samples and establish a contrast loss. As shown in ③ of Figure 1 , the present invention uses the known anchor links as positive samples, and randomly constructs a certain proportion of negative samples for each positive sample, and constructs a contrast loss by comparing the distances between the positive samples and the negative samples in the vector representation space;

[0060] Step 4: Construct the overall optimization objective of the algorithm and joint optimization of the algorithm. Use the weighted combination of the reconstruction loss and the contrast loss to construct the overall optimization objective of the proposed algorithm, and realize the joint optimization of the algorithm through the error backpropagation algorithm to obtain the optimized spatial consistency representation of the nodes;

[0061] Step 5: Inference of unknown anchor links (i.e., cross-social network user identity recognition). For a node in the source (target) network , based on the optimized spatial consistency representation of the nodes, the present invention uses the Euclidean distance as the metric method to find the k nodes with the closest distance to it in the target (source) network as the candidate set, as shown in ④ of Figure 1 ;

[0062] Among them, steps 1, 2, 3, and 4 are the key steps of the proposed algorithm, and the present invention will elaborate in detail next.

[0063] 1.2 Node Feature Representation Based on Graph Convolutional Neural Network

[0064] As Figure 1 shown, the present invention uses the adjacency matrices A s and A t of the source network and the target network as the inputs of the encoder, and realizes the transmission and aggregation of node information in the social network through the graph convolutional layer:

[0065]

[0066] Where and respectively represent the input and output of the (l + 1)-th layer of GCN, d (l) represents the vector dimension of the l-th layer, I n is the identity matrix, represents the degree matrix of, is the weight matrix of the (l + 1)-th layer of GCN, and σ represents the activation function.

[0067] Thus, the present invention can obtain the feature representations of the nodes in the source network and the target network:

[0068]

[0069]

[0070] Where U s , U t respectively represent the node representations output by the encoders of the source network and the target network, and are in the same vector representation space; are respectively the initial feature representations of the nodes in the source network and the target network. Since the present invention only conducts relevant research in the scenario where the social network relationship structure is available, the present invention randomly initializes the initial feature representations of the nodes.

[0071] 1.3 Reconstruct the Original Network Structure and Establish the Reconstruction Loss

[0072] Reconstruct the adjacency matrix. As Figure 1 shown in ② of, based on the node feature representations obtained in the first step, the present invention reconstructs the adjacency matrix of the source network using the inner product decoder:

[0073]

[0074] Where (U s ) T represents the transpose of U s ; σ is the activation function Sigmoid; Represents the reconstructed adjacency matrix of the source network. Similarly, the present invention can also obtain the reconstructed adjacency matrix of the target network:

[0075]

[0076] Construct the reconstruction loss. Construct the reconstruction loss through the difference between the reconstructed adjacency matrix and the original adjacency matrix:

[0077]

[0078]

[0079] where y s and y t represent the elements in the adjacency matrix A s of the source network and the adjacency matrix A t of the target network respectively, and their values are 0 or 1; and represent the elements in the reconstructed adjacency matrix of the source network and the reconstructed adjacency matrix of the target network respectively; N s and N t represent the number of nodes in the source network and the target network respectively.

[0080] 1.4 Construct positive and negative samples and establish the contrast loss

[0081] Construct positive and negative samples. For any positive sample The present invention randomly searches for K nodes in the target network for the anchor node so as to construct K negative samples Similarly, the present invention can search for K nodes in the source network for the anchor node so as to construct K negative samples again The present invention refers to the set of positive and negative samples as the training sample set, denoted as

[0082] Establish the contrast loss. For any sample in the training sample set The contrast loss is expressed as follows:

[0083]

[0084] where y label represents the label of the sample , the positive sample is 1, and the negative sample is 0; represents the Euclidean distance of the sample ; represents the number of samples in the training sample set; margin is the set threshold.

[0085] 1.5 Construction of Loss Function and Joint Optimization of Algorithm

[0086] Overall Optimization Objective. Combining formulas (6), (7), and (8), the present invention constructs the overall loss function of algorithm CLink in a weighted combination manner:

[0087]

[0088] where α is the weight coefficient that controls the proportion of reconstruction loss and contrast loss.

[0089] Joint Optimization of Algorithm. The present invention uses the Adam optimizer to minimize the overall loss function (9) and applies the error backpropagation algorithm to achieve iterative optimization of the proposed algorithm. Specifically, by minimizing the reconstruction loss, the learned node representations are made as close as possible to the original structural features of the source network and the target network, ensuring that nodes that are structurally similar are still similar in the representation space. By minimizing the contrast loss, the positive samples are made closer and the negative samples are made farther apart in the vector representation space, making the learned node representations more discriminative.

[0090] 2. Experiments

[0091] To verify the effectiveness of the method proposed in the present invention in the problem of cross-social network user identity recognition, the present invention conducts experiments on three real social network datasets respectively.

[0092] 2.1 Experimental Settings

[0093] 2.1.1 Experimental Datasets

[0094] The present invention conducts experiments on three publicly available real social network datasets, and the detailed statistical data of the datasets are shown in Table 1.

[0095] Dataset 1 (Douban-Weibo): Provided by the Cao Xuezhi team of Shanghai Jiao Tong University [Cao X, Yu Y. ASNets: A benchmark dataset of aligned social networks for cross-platform user modeling [C] / / Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. 2016: 1881-1884.], containing two social networks, Douban and Weibo, which can be obtained online (https: / / apex.sjtu.edu.cn / datasets / 1);

[0096] Dataset 2 (Foursquare-Twitter): Provided by the team of Philip S. Yu from the University of Illinois at Chicago, USA [Kong X, Zhang J, Yu P S. Inferring anchor links across multiple heterogeneous social networks [C] / / Proceedings of the 22nd ACM international conference on Information & Knowledge Management. 2013:179-188.]. It contains two social networks, Foursquare and Twitter, and can be obtained from GitHub (https: / / github.com / ColaLL / IONE);

[0097] Dataset 3 (Dblp17-Dblp19): Provided by the paper MAUIL [Chen B, Chen X. MAUIL: Multilevel attribute embedding for semi-supervised user identity linkage [J]. Information Sciences, 2022, 593:527-545.]. It can be obtained from GitHub (https: / / github.com / ChenBaiyang / MAUIL). The authors of the paper selected the Turing Award winner Yoshua Bengio as the central node in each network, and then deleted all nodes more than three hops away from the center, obtaining two networks, Dblp17 and Dblp19.

[0098] Table 1 Statistical details of the datasets

[0099]

[0100] 2.1.2 Baseline methods

[0101] The proposed algorithm CLink of the present invention is experimentally compared and analyzed with six state-of-the-art baseline algorithms based on the social network relationship structure. The detailed descriptions of the baseline algorithms are as follows:

[0102] PALE (Reference [1]: Man T, Shen H, Liu S, et al. Predict anchor links across social networks via an embedding approach[C] / / Ijcai. 2016, 16: 1823 - 1829.): The PALE model applies network representation learning (e.g., DeepWalk, LINE, etc.), uses the known anchor links as supervision information, and uses a multi-layer perceptron to learn a mapping function across social networks for anchor link prediction. When reproducing the PALE model in this paper, DeepWalk is used to learn the vector representation of nodes;

[0103] IONE (Reference [2]: Chu X, Fan X, Zhu Z, et al. Variational cross-network embedding for anonymized user identity linkage[C] / / Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 2021: 2955 - 2959.): The IONE model combines network representation learning based on second-order proximity and the learning of network alignment for optimization in a unified framework;

[0104] GraphUIL (Reference [3]: Zhang W, Shu K, Liu H, et al. Graph neural networks for user identity linkage[J]. arXiv preprint arXiv:1903.02174, 2019.): The GraphUIL model is an alignment method based on graph neural networks. It uses two aggregators to learn global structural features and local structural features respectively, and uses a multi-layer perceptron model to learn the mapping function between the embedding spaces of two social networks;

[0105] DANA (Reference [4]: Hong H, Li X, Pan Y, et al. Domain-adversarial network alignment [J]. IEEE Transactions on Knowledge and Data Engineering, 2020: 3211-3224.): The DANA model is a domain-adversarial based alignment method. First, it uses a graph convolutional network to learn the low-dimensional vector representation of nodes under the principle of domain adversariality, and then maximizes the posterior probability of anchor nodes and minimizes the domain classifier loss in a unified framework for optimization;

[0106] DANA-S (Reference [4]): DANA-S is a variant of the DANA model. Different from the DANA model, DANA-S uses a graph convolutional network with shared weights to learn the low-dimensional vector representations of the source network and the target network;

[0107] BRIGHT (Reference [5]: Yan Y, Zhang S, Tong H. Bright: A bridging algorithm for network alignment [C] / / Proceedings of the Web Conference 2021. 2021: 3907-3917.): The BRIGHT model takes the one-hot encoded vector of the training set anchor links as the basis and constructs a unified vector representation space through restart random walk.

[0108] 2.1.3 Evaluation Metrics

[0109] The present invention uses the most common precision@k metric to evaluate the performance of the algorithm proposed in the present invention and the benchmark methods, which is defined as follows:

[0110]

[0111] where S test represents the test set, and |S test | is the number of anchor links in the test set; for the anchor link the present invention finds the k most similar nodes to node in all nodes of the target network (source network) as its top-k candidate set; success i @k represents whether appears in the top-k candidate set of i , if so, success i @k = 1, otherwise success j @k represents Whether it appears in the top-k candidate set.

[0112] In addition to using the precision@k metric, the present invention also uses the Mean Average Precision (MAP) to evaluate the model performance, which is defined as follows:

[0113]

[0114] For the anchor link the present invention sorts the similarities between the node and all nodes in the target network (source network) in descending order as the candidate set, where ra i represents the position where it appears in the candidate set, and ra j represents the position where it appears in the candidate set.

[0115] It should be noted that the method proposed by the present invention and all the baseline algorithms reproduced in this embodiment are evaluated according to the calculation methods defined by formulas (13) and (14). For example, the baseline method BRIGHT uses a one-way calculation method from the source network to the target network, and the present invention has uniformly changed it to a two-way calculation method from the source network to the target network and from the target network to the source network.

[0116] 2.1.4 Experimental Parameter Settings

[0117] In the experiments of the present invention, 80% of all known anchor links are randomly selected for model training, and the remaining anchor links are used for model testing. To avoid the influence of experimental randomness and contingency factors on the experimental results, the present invention randomly selects each dataset 5 times and conducts experiments independently in sequence, and compares the average values of the results. The present invention randomly initializes the initial features of the nodes, with the dimension set to 500; sets 2 graph convolutional layers with shared weights, and the dimensions are 256 and 100 respectively; the negative sample ratio K is set to 20; the model optimizer uses the Adam optimizer, and the learning rate is set to 0.01; the batch size is set to 512; the weight coefficient alpha of the reconstruction loss and the contrast loss ratio is set to 0.5.

[0118] For the fairness of experimental comparison, the negative sample number K is set to 20 when constructing negative samples in the baseline methods, which is the same as the setting of the present invention method. All baseline methods adopt the strategy of taking the average value of multiple experiments. Other hyperparameters of the baseline methods are set according to the corresponding articles. All the experiments in the paper are carried out on an AMAX server, and the server CPU is Xeon(R) Silver 4210 CPU @ 2.20GHz, the GPU is NVIDIA Corporation TU102, and the memory is 128G. The algorithm proposed by the present invention is implemented using the PyTorch framework.

[0119] 2.2 Experimental Results

[0120] The present invention compares the performance of CLink and other benchmark methods on three real social network datasets. Table 2 shows the average values of the experimental results obtained from multiple experiments under the evaluation metrics precision@1, precision@5, precision@10, and MAP, where 80% of the anchor links are randomly selected as the training set, and precision@k is abbreviated as p@k. For the convenience of comparative analysis, the present invention highlights the best results under each metric in bold. The specific experimental results of all methods at different precision@k on the three datasets are as Figure 2 shown. The experimental results of the MAP metric are as Figure 3 shown, and the result data of the two best-performing methods on different datasets are shown above the bar chart.

[0121] Table 2 Average performance of all methods on three datasets (training ratio = 0.8)

[0122]

[0123] (1) Overall performance analysis. As can be seen from Table 2 and Figure 2 , except for the precision@1 metric on the Dblp17 - Dblp19 dataset, the performance of the CLink method of the present invention is better than that of all benchmark methods. For the evaluation metrics precision@1 and precision@5, CLink improves by 6.5% and 9.3% respectively compared to the best-performing benchmark method GraphUIL on the Douban - Weibo dataset; on the Foursquare - Twitter dataset, it improves by 2.3% and 3.4% respectively compared to the best benchmark method DANA - S; on the Dblp17 - Dblp19 dataset, although the precision@1 performance of the benchmark method DANA - S is 1.3% higher than that of the method of the present invention, in terms of precision@5 and precision@10, the performance of the method of the present invention is significantly higher than that of DANA - S, with improvements of 5.7% and 7.5% respectively.

[0124] (2) From Table 2 and Figure 2As can be seen, the PALE model is one of the benchmark methods with the worst performance, with a precision@10 of only about 7% on the Douban-Weibo dataset, while the model of the present invention can reach more than 32%. Among all the benchmark methods, PALE is a two-stage model. Learning the vector representation of nodes and identifying user identities are two independent processes, which cannot guarantee that the learned node vector representation is beneficial to the downstream user identity recognition task, resulting in poor model performance.

[0125] (3) As can be seen from Table 2, on the Douban-Weibo dataset, GraphUIL is the benchmark method with the best performance, and the precision@10 can reach more than 23%. From Figure 2 (b) It can be seen that on the Foursquare-Twitter dataset, when k is small, the performance of the GraphUIL method is poor. However, when k reaches 70, GraphUIL is comparable to the better benchmark methods DANA and IONE. A similar performance is shown on the Dblp17-Dblp19 data. When k reaches about 20, it is comparable to the performance of the benchmark method DANA-S. The GraphUIL method uses a multi-layer perceptron to learn the non-linear mapping function between the source network and the target network, reducing the distance between positive samples with corresponding relationships. Compared with the method CLink of the present invention, GraphUIL uses independent graph neural networks to learn the vector representations of nodes in the source network and the target network, cannot map nodes to a unified vector representation space, and lacks the consideration of the influence of negative samples on the user identity recognition problem. It can also be seen from the experimental results that on the three datasets, the performance of the method of the present invention is better than that of GraphUIL.

[0126] (4) It can be seen from the experimental results that the DANA model and its variant DANA-S are one of the benchmark methods with the best performance. Generally speaking, DANA and DANA-S perform better on the Foursquare-Twitter dataset and the Dblp17-Dblp19 dataset. On the Foursquare-Twitter dataset, the precision@1 and MAP metrics of the DANA-S model are only 2.3% and 2.7% lower than those of the method CLink of the present invention respectively. On the Dblp17-Dblp19 dataset, the precision@1 metric is even 1.3% higher than that of the method of the present invention. However, in the metrics after Precision@5, the method of the present invention shows better performance. DANA-S performs poorly on the douban-weibo dataset, with a difference of 7.5% and 8.7% from the method of the present invention in the precision@1 and MAP metrics respectively.

[0127] Different from the baseline method GraphUIL, DANA and DANA-S model the relative distances between positive and negative samples, and at the same time design a domain classifier based on adversarial learning to filter out domain-related features and focus on extracting alignment-related features. The differences from the method of the present invention lie in the reconstruction loss and the domain classifier loss. From the experimental results, generally the method of the present invention shows better performance.

[0128] (5) As can be seen from Table 2 and Figure 2 it can be seen that when k is less than 10, IONE and BRIGHT have similar performance, and generally the experimental results of BRIGHT are slightly higher than those of IONE. However, when k is greater than 10, the performance improvement of BRIGHT is slow, and the effect after precision@10 is significantly lower than that of IONE, DANA and the method of the present invention.

[0129] Compared with GraphUIL, DANA and the method of the present invention, the two methods of IONE and BRIGHT do not apply graph neural networks to learn the vector representations of nodes, and it can be seen that reasonably applying graph neural networks can improve the accuracy of user identity recognition to a certain extent. However, compared with the two-stage PALE, IONE and BRIGHT optimize the vector representation based on random walk and the downstream user identity recognition task in a unified framework, showing better performance than PALE.

[0130] 2.3 Experimental parameter analysis

[0131] In this section, the present invention first explores the influence of the anchor link ratio (training ratio) and the number of negative samples in the training set on the experimental results. When analyzing the influence of one factor, the other factor is kept unchanged. Then the present invention conducts a series of ablation experiments to explore the influence of reconstructing the original network structure and the positive and negative sample comparison strategy on the experimental results. All experiments in this subsection are carried out on the Foursquare-Twitter dataset, and the model is evaluated using the precision@1 and MAP metrics.

[0132] 2.3.1 Influence of the anchor link ratio in the training set on the experimental results

[0133] To explore the influence of the anchor link ratio in the training set on the experimental results, the present invention fixes the number of negative samples K at 20, sets the anchor link ratio in the training set to 0.1 - 0.9, with an interval of 0.1 each time, and conducts 9 independent experiments. Figure 4 Shows the average values of multiple experimental results under different training set ratios. From Figure 4As can be seen, as the proportion of the training set increases, the performance of all methods is continuously improving. When the training ratio is less than 0.5, IONE performs the worst among all methods; when the training ratio reaches more than 0.5, the GraphUIL method grows slowly and performs the worst. For the precision@1 metric, the three methods DANA, DANA-S, and BRIGHT perform equivalently, but for the MAP metric, DANA and DANA-S perform slightly better than BRIGHT. From Figure 4 As can be seen from (a) and (b), under 9 different training set ratios, the method proposed by the present invention is superior to all baseline methods, and also has good performance when the training ratio is relatively low. For example, when the training ratio reaches 0.3, the precision@1 and MAP of the method of the present invention can reach 9% and 15% respectively.

[0134] 2.3.2 Influence of the number of negative samples on the experimental results

[0135] To explore the influence of the number of negative samples K on the experimental results, the present invention fixes the training set ratio at 0.8 and sets the number of negative samples to 1, 5, 10, 20, 30, 40, 50, and 100 respectively, and conducts 8 independent experiments. Figure 5 Shows the average values of the experimental results under different numbers of negative samples. From Figure 5 As can be seen, when the number of negative samples reaches about 20, the performance of all methods gradually levels off. From the curves of the performance changes of DANA and DANA-S, it can be seen that these two baseline methods are greatly affected by the number of negative samples. When the number of negative samples is 1, the precision@1 value of DANA-S is less than 7%, and the precision@1 of DANA is only about 5%. Compared with DANA and DANA-S, the baseline method BRIGHT and the method of the present invention still have excellent performance when the number of negative samples is small. For example, when only one negative sample is extracted, the performance of the BRIGHT method is almost the same as that of the method of the present invention. Generally speaking, under 8 different negative sample conditions, the method of the present invention has better performance than the baseline methods.

[0136] 2.3.3 Ablation experiments

[0137] To verify the rationality of the present invention in jointly optimizing the node consistency feature representation using the reconstruction loss and the contrastive loss, the present invention conducts a series of ablation experiments, and the specific settings are as follows:

[0138] ·CLink: The complete algorithm CLink proposed by the present invention;

[0139] · Ablation-1: Remove the negative sample constraint from CLink. In the unified vector representation space, the present invention uses the mean squared error loss function (MSELoss) to reduce the distance between positive samples in the vector representation space (reconstruct the original network structure + only use positive samples);

[0140] · Ablation-2: Remove the reconstruction process from CLink. Only use the contrastive loss as the final loss of the overall framework (positive and negative sample contrast);

[0141] · Ablation-3: Remove the reconstruction process and negative sample constraint from CLink. Only use the mean squared error loss as the final loss of the overall model (only use positive samples).

[0142] The present invention conducts three ablation experiments on the Foursquare-Twitter dataset with a training set ratio of 0.8, and other hyperparameter settings remain the same as those of CLink. The experimental results are shown in Table 3.

[0143] · Clink vs Ablation-1 and Ablation-2 vs Ablation-3: As can be seen from Table 3, the precision@1 of the Ablation-1 model reaches more than 14%, and the MAP value also exceeds 21%. The overall performance of the Ablation-1 model is relatively close to the best DANA-S method among the baseline methods, but there is still a certain gap compared with the complete algorithm CLink proposed by the present invention. Through the comparison of the results of these two groups of experiments, the present invention can see the effectiveness of the positive and negative sample contrast strategy in solving the user identity recognition problem.

[0144] · Clink vs Ablation-2 and Ablation-1 vs Ablation-3: Compared with CLink, the accuracy of the Ablation-2 model without the reconstruction process has a relatively obvious decline. Similarly, compared with Ablation-1, the accuracy of the Ablation-3 model also shows the same decline phenomenon. Through the analysis of the results of these two groups of experiments, the present invention can see that reconstructing the original network structure has a greater impact on the user identity recognition problem.

[0145] · Ablation-2 vs GraphUIL: As can be seen from the comparison, the Ablation-2 model outperforms the baseline algorithm GraphUIL model by 4.6% and 6.2% respectively in terms of precision@1 and MAP metrics. Both models include a reconstruction process and positive sample constraints. The difference between them is that GraphUIL learns the feature representations of the source network and the target network separately, while Ablation-2 encodes different social networks into a unified vector representation space. The experimental results further illustrate the effectiveness of maintaining the consistency of the node representation space in the problem of user identity recognition.

[0146] The setting of the ablation experiment and the comparative analysis of the results of 5 groups of comparative experiments illustrate, on the one hand, the effectiveness of encoding the source network and the target network into a consistent vector space, and on the other hand, the effectiveness and rationality of using the reconstruction process and the joint constraints of positive and negative samples in the node representation process of the present invention in solving the problem of user identity recognition.

[0147] Table 3 Results of the ablation experiment on the Foursquare-Twitter dataset (training ratio = 0.8)

[0148] Evaluation Metrics precision@1 precision@5 precision@10 MAP CLink 0.17050 0.32578 0.40839 0.24765 Ablation-1 0.14193 0.29317 0.37391 0.21694 Ablation-2 0.06522 0.13354 0.19255 0.10703 Ablation-3 0.00248 0.00466 0.00776 0.00595

[0149] Based on the above embodiments, as Figure 6 shown, the present invention also proposes a cross-social network user identity recognition system based on spatially consistent representation, including:

[0150] A node consistency representation module, which is used to encode the source network and the target network from the structural space into a unified vector representation space by stacking multiple graph convolutional layers as an encoder, and obtain the spatially consistent representation of user nodes; both the source network and the target network are social networks;

[0151] A network structure reconstruction module, which is used to reconstruct the network structures of the source network and the target network by using an inner product decoder, and construct a reconstruction loss by minimizing the difference between the reconstructed network and the original network;

[0152] A contrast loss construction module, which is used to use known anchor links as positive samples, and randomly construct a certain proportion of negative samples for each positive sample, and construct a contrast loss by the constraint that the distance of positive samples is closer than that of negative samples in the vector representation space;

[0153] A joint optimization module, which is used to optimize the spatially consistent representation of nodes by the joint constraints of the reconstruction loss and the contrast loss;

[0154] A user identity recognition module, which is used to perform cross-social network user identity recognition based on the optimized spatially consistent representation of nodes.

[0155] Furthermore, the expression of the reconstruction loss is as follows:

[0156]

[0157]

[0158] where y s and y t represent the elements in the source network adjacency matrix A s and the target network adjacency matrix A t respectively, and their values are 0 or 1; and represent the elements in the source network reconstructed adjacency matrix and the target network reconstructed adjacency matrix respectively; N s and N t represent the number of nodes in the source network and the target network respectively; and U s and U t represent the node representations output by the encoders of the source network and the target network respectively, (U s ) T and (U t ) T represent the transposes of U s and U t respectively, and σ is the activation function Sigmoid.

[0159] Furthermore, the contrastive loss is:

[0160]

[0161] where y label represents the label of the sample , the positive sample is 1, and the negative sample is 0; represents the Euclidean distance of the sample ; represents the number of samples in the training sample set; margin is the set threshold.

[0162] Furthermore, the joint optimization module is specifically used for:

[0163] Combining formulas (6), (7), and (8), constructing an overall loss function in a weighted combination manner:

[0164]

[0165] where α is the weight coefficient controlling the ratio of the reconstruction loss and the contrastive loss;

[0166] The Adam optimizer is used to minimize the overall loss function (9), and the error backpropagation algorithm is applied to achieve iterative optimization, obtaining a spatially consistent representation based on the optimized nodes.

[0167] In summary, the present invention studies the problem of cross-social network user identity recognition, models learning user feature representation and user identity recognition in a unified framework, and proposes a cross-social network user identity recognition method and system based on spatially consistent representation. The present invention uses a graph convolutional neural network to encode the source network and the target network into a unified vector representation space, maintaining the spatial consistency of node representations. At the same time, two parts of constraints, namely reconstruction loss and contrast loss, are used to jointly optimize the representation process of nodes, that is, it not only ensures that the learned node representations are as close as possible to the original structural features of the source network and the target network, but also ensures that in the unified vector representation space, the anchor node pairs are closer than the non-anchor node pairs. Experimental results on three real social network datasets show that the performance of the method proposed by the present invention is superior to all baseline methods. At the same time, the results of ablation experiments further illustrate the rationality and effectiveness of the method proposed by the present invention.

[0168] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for cross - social network user identity recognition based on spatial consistency representation, characterized in that, Including: Stacking multiple graph convolutional layers as an encoder to encode the source network and the target network from the structural space into a unified vector representation space, and obtaining a spatially consistent representation of user nodes; both the source network and the target network are social networks; Using an inner product decoder to reconstruct the network structures of the source network and the target network, and constructing a reconstruction loss by minimizing the difference between the reconstructed network and the original network; Taking known anchor links as positive samples, and randomly constructing a certain proportion of negative samples for each positive sample, and constructing a contrastive loss by the constraint that the distance of positive samples is closer than that of negative samples in the vector representation space; Using the joint constraints of the reconstruction loss and the contrastive loss to optimize the spatially consistent representation of nodes; Performing cross-social network user identity recognition based on the optimized spatially consistent representation of nodes.

2. The cross-social network user identity recognition method based on spatial consistency representation according to claim 1, wherein The expression of the reconstruction loss is: where y s and y t represent the elements in the source network adjacency matrix A s and the target network adjacency matrix A t respectively, and their values are 0 or 1; and represent the elements in the source network reconstructed adjacency matrix and the target network reconstructed adjacency matrix respectively; N s and N t represent the number of nodes in the source network and the target network respectively; and U s 、U t represent the node representations output by the encoders of the source network and the target network respectively, (U s ) T 、(U t ) T represent the transposes of U s 、U t respectively, and σ is the activation function Sigmoid.

3. The cross-social network user identity recognition method based on spatial consistency representation according to claim 2, wherein The contrastive loss is: where y label represents the label of the sample , the positive sample is 1 and the negative sample is 0; represents the Euclidean distance of the sample ; represents the number of samples in the training sample set; margin is the set threshold.

4. The cross-social network user identity recognition method based on spatial consistency representation according to claim 3, wherein The using the joint constraints of the reconstruction loss and the contrastive loss to optimize the spatially consistent representation of nodes includes: Combining formulas (6), (7), and (8), and constructing an overall loss function by a weighted combination method: where α is a weight coefficient for controlling the ratio of the reconstruction loss and the contrastive loss; Using an Adam optimizer to minimize the overall loss function (9), and applying the error backpropagation algorithm to achieve iterative optimization, and obtaining the spatially consistent representation of nodes based on the optimization.

5. A cross-social network user identity recognition system based on spatial consistency representation, characterized in that, Including: A node consistency representation module, which is used to stack multiple graph convolutional layers as an encoder to encode the source network and the target network from the structural space into a unified vector representation space, and obtain a spatially consistent representation of user nodes; both the source network and the target network are social networks; A network structure reconstruction module, which is used to use an inner product decoder to reconstruct the network structures of the source network and the target network, and construct a reconstruction loss by minimizing the difference between the reconstructed network and the original network; A contrastive loss construction module, which is used to take known anchor links as positive samples, and randomly construct a certain proportion of negative samples for each positive sample, and construct a contrastive loss by the constraint that the distance of positive samples is closer than that of negative samples in the vector representation space; A joint optimization module, which is used to use the joint constraints of the reconstruction loss and the contrastive loss to optimize the spatially consistent representation of nodes; A user identity recognition module, which is used to perform cross-social network user identity recognition based on the optimized spatially consistent representation of nodes.

6. The cross-social network user identity recognition system based on spatial consistency representation according to claim 5, characterized in that The expression of the reconstruction loss is: where y s and y t represent the elements in the source network adjacency matrix A s and the target network adjacency matrix A t respectively, and their values are 0 or 1; and represent the elements in the source network reconstructed adjacency matrix and the target network reconstructed adjacency matrix respectively; N s and N t represent the number of nodes in the source network and the target network respectively; and U s and U t represent the node representations output by the encoders of the source network and the target network respectively. (U s ) T and (U t ) T represent the transposes of U s and U t respectively, and σ is the activation function Sigmoid.

7. The cross-social network user identity recognition system based on spatial consistency representation according to claim 6, wherein The contrastive loss is: where y label represents the label of the sample , with the positive sample being 1 and the negative sample being 0; represents the Euclidean distance of the sample ; represents the number of samples in the training sample set; margin is the set threshold.

8. The cross-social network user identity recognition system based on spatial consistency representation according to claim 7, characterized in that, The joint optimization module is specifically used for: Combining formulas (6), (7), and (8), and constructing an overall loss function by a weighted combination method: where α is a weight coefficient for controlling the ratio of the reconstruction loss and the contrastive loss; Using an Adam optimizer to minimize the overall loss function (9), and applying the error backpropagation algorithm to achieve iterative optimization, and obtaining the spatially consistent representation of nodes based on the optimization.

Images