Cross-social network account matching method based on knowledge graph label propagation

Abstract

The invention discloses a cross-social network account matching method based on knowledge graph label propagation, and the method comprises the following steps: S1, obtaining user account data on two social network platforms, and constructing a user knowledge graph of the user account data; s2, performing false label propagation based on the degrees of the nodes in the knowledge graph; and S3, label propagation based on user information similarity: taking the'to-be-matched node 'as a center, firstly positioning the'bridging node' of the'to-be-matched node ', then constructing a potential matching node set by adopting a'first-order neighbor union set' through a matching relationship of the bridging node, and completing label updating and propagation based on the user information similarity. According to the method, different social network user data are directly processed and associated from the perspective of the graph, and the accuracy of cross-platform social network account matching is remarkably improved by deeply mining the topological relation and semantic information of the nodes and the connecting edges in the knowledge graph.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of multi-source heterogeneous data correlation analysis, and particularly relates to a cross-social network account matching method based on knowledge graph label propagation. BACKGROUND

[0002] In the current high-speed development of information technology, data is growing explosively, and the sources are increasingly diverse, the structure and semantics are increasingly complex, and multi-source heterogeneous data exists widely in various fields. Multi-source heterogeneous data refers to data from different data sources, with different data structures, storage formats and semantic meanings, such as different types of text, images, and data from different databases and systems. Multi-source heterogeneous data correlation aggregates data with different sources, structures and semantics, for example, in the field of social networks, user data on different platforms such as Twitter and LinkedIn needs to be analyzed to achieve user portrait integration, cross-platform behavior analysis and other operations. The correlation analysis of these multi-source heterogeneous data is a key link to obtain valuable information from them.

[0003] Due to the differences in data format, quality, semantic understanding, etc., it is extremely challenging to realize the association of multi-source heterogeneous data. Traditional attribute-based matching methods (Elmagarmid AK, Ipeirotis P G, Verykios V S. Duplicate record detection: A survey [J]. IEEE Transactions on knowledge and data engineering, 2006, 19(1): 1-16) rely on comparing the key attributes of entities, such as names, identifiers, etc. for association, but the effect is not good when facing complex data. Machine learning-based methods (Christen P. A survey of indexing techniques for scalable record linkage and deduplication [J]. IEEE transactions on knowledge and data engineering, 2011, 24(9): 1537-1555) use classification and clustering algorithms to mine the potential relationships of data, but they need a large amount of labeled data to train the model. Emerging graph algorithms, like graph neural networks (Sun Z, Deng Z H, Nie J Y, et al. Rotate: Knowledge graph embedding by relational rotation in complex space [J]. arxiv preprint arxiv:1902.10197, 2019), knowledge graphs, realize association by mining the data features of graph data structure, and show great potential.

[0004] Knowledge graphs (Fensel D, U, Angele K, et al. Introduction: what is a knowledge graph? [J]. Knowledge graphs: Methodology, tools and selected use cases, 2020: 1-10) construction is a key technology for integrating multi-source data and forming a structured semantic network. The process involves multiple steps. In the data collection stage, various types of data such as text, database records, multimedia information, etc. are widely collected to provide rich materials for subsequent processing.

[0005] However, the existing multi-source heterogeneous data analysis matching method faces many challenges in processing these data. On the one hand, the traditional method highly depends on manual feature engineering, which requires professional personnel to spend a lot of time and effort to design and extract data features, and is inefficient and difficult to adapt to complex and variable data environment. On the other hand, due to the heterogeneity of data, the data structure and semantic difference of different data sources are large, which leads to information fragmentation and makes it difficult to capture the complex correlation between data. At the same time, when constructing a knowledge graph for data correlation, there are problems of different knowledge graph density and less known information, which makes the original information unable to effectively spread outward, seriously affecting the efficiency and accuracy of multi-source heterogeneous data correlation matching, and cannot meet the demand of data depth analysis and efficient processing in practical application.

[0006] Therefore, how to efficiently and accurately realize the correlation analysis of data of users of different platforms has become a problem to be solved at present. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art and provide a cross-social network account matching method based on knowledge graph label propagation. The method of the present application can make full use of the attribute information of nodes and edges in the knowledge graph, intuitively process and correlate multi-source heterogeneous data, reduce the dependence on manual feature engineering, effectively capture complex correlation, and break the dilemma of information fragmentation. Moreover, through the innovative method of constructing false labels, the problem of label propagation caused by sparse knowledge graph and few known matching pairs is solved, the efficiency and accuracy of account correlation matching are improved, and a more effective solution is provided for data processing and analysis in related fields.

[0008] The purpose of the present application is achieved by the following technical scheme: a cross-social network account matching method based on knowledge graph label propagation, comprising the following steps:

[0009] S1, respectively acquiring user account data on two social network platforms, and constructing user knowledge graphs of each user account data; the specific steps are as follows:

[0010] S1-1, taking each object in the two social network platforms as a node, adding a label "Label" to each node, the label being a node on another social network platform, and the initial value being None, which is iterated through subsequent continuous propagation; the label value is "Score", representing the similarity between the node and the label, and the initial value is 0, ranging from 0 to 1;

[0011] and adding an "Info" field for storing user information; the nodes in one knowledge graph are named t_*, and the nodes in the other knowledge graph are named l_*;

[0012] S1-2. Construct undirected edges in the knowledge graphs of the two user accounts according to their respective social relationships, representing that the two are friends.

[0013] S2. A method for propagating false labels based on the degree of nodes in a knowledge graph: Define the degree of a node in a knowledge graph as the number of edges directly connected to that node; set a threshold for degree, and designate nodes with a degree greater than the threshold as "socially active" nodes; arrange the "socially active" nodes in two knowledge graphs in order of degree, and then match the "socially active" nodes in the two knowledge graphs respectively. Nodes with a similarity exceeding the threshold between the "socially active" nodes in the two knowledge graphs are matched and designated as the correct labels for the other node; and set false labels for these matched "socially active" nodes.

[0014] S3. Tag propagation method based on user information similarity: In the input data, a small number of nodes already have correct labels. These nodes propagate outwards through false labels, allowing more unlabeled nodes in the two knowledge graphs to find their matching nodes. The specific implementation steps are as follows:

[0015] S3-1. Sort the nodes in the two knowledge graphs according to their degree.

[0016] S3-2. In order of degree, traverse any node in the knowledge graph from largest to smallest. Each traversed node is called a node to be matched. Find the nodes with the correct labels among the first-order neighbors of the node to be matched. These nodes are called the "bridging nodes" of the node to be matched. The set of "bridging nodes" is called the "bridging node set" of the node to be matched. The first-order neighbor is defined as the node directly connected to the node to be matched.

[0017] S3-3. By analyzing the "bridging nodes" of the node to be matched, determine the potential matching node set of the node to be matched. Specifically, based on the "bridging nodes" of the node to be matched, find the nodes in another knowledge graph that match these "bridging nodes", and take the union of the first-order neighbors of these nodes as the potential matching node set of the node to be matched in the other knowledge graph.

[0018] S3-4. Calculate the similarity between two nodes by using the basic user information stored in the "Info" field of the node to be matched and the node in the set of potential matching nodes. Record the node with the highest similarity, which is called the "most similar node" of the node to be matched.

[0019] If the similarity of the record is greater than the existing "Score" label value of the "most similar node", the existing "Label" label of the "most similar node" is updated to point to the node to be matched, and the "Score" value is updated to the highest similarity; and the "Label" label of the corresponding node previously pointing to the current "most similar node" is reset to "null", and the "Score" is set to 0;

[0020] If the similarity of the current node to be matched and its "most similar node" is less than the existing "Score" label value of the "most similar node", the "most similar node" of the node to be matched is updated to point to the node with the second largest similarity in the potential matching node set, and the comparison process is repeated.

[0021] S3-5, after each iteration, a new matching pair appears, and the number of matching pairs generated in each round is recorded, and when the number of matching pairs is 0, it indicates that no new node pair can be matched, the loop is exited, the iteration is stopped, and the label propagation is ended.

[0022] The beneficial effects of the present application are: the present application proposes a cross-social network account matching method based on knowledge graph label propagation, which directly processes and associates different social network user data from the perspective of graph, and also proposes a method of constructing false labels to solve the problem that the original label cannot be propagated outward due to the sparsity of the knowledge graph and the small number of known matching pairs, which can finally bring the following beneficial effects:

[0023] 1. The present application is based on knowledge graph technology and innovatively proposes a social network user data processing and association method. By deeply mining the topological relationship and semantic information of nodes and edges in the knowledge graph, the accuracy of cross-platform social network account matching is significantly improved. This method effectively solves the semantic alignment and entity association problem of multi-source heterogeneous user data in the social network environment, and provides a new technical path for cross-social media platform account matching.

[0024] 2. The present application creatively proposes a knowledge graph association method based on false label propagation. By intelligently selecting the key nodes with the highest degree centrality in the knowledge graph to establish cross-graph association, the label information is efficiently propagated in multiple paths while ensuring the matching accuracy. This innovation not only solves the low matching efficiency problem caused by the sparsity of initial information in traditional methods, but also effectively balances the structural heterogeneity between different knowledge graphs, significantly improves the alignment effect between graphs with large density differences, and improves the overall matching efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The flowchart of the cross-social network account matching method based on knowledge graph label propagation of the present application;

[0026] Figure 2 Construct an example graph for the set of potential matching nodes. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, the present invention provides a cross-social network account matching method based on knowledge graph tag propagation, comprising the following steps:

[0029] S1. Acquire user account data from two social networking platforms. This embodiment uses Twitter and LinkedIn as examples to construct user knowledge graphs for each user account. This step is to convert user data from different platforms (such as basic user information and social relationships) into two independent knowledge graphs: a Twitter user knowledge graph and a LinkedIn user knowledge graph. This allows the application of relationship reasoning and tag propagation algorithms on the knowledge graphs for subsequent cross-platform user matching relationship reasoning. The specific steps are as follows:

[0030] S1-1. Treat each object in the two social networking platforms, Twitter and LinkedIn, as a node. Add a label "Label" to each node, where the label represents the node on the other social networking platform. The initial value of the label is None, and it is increased through continuous propagation and iteration. The label value is "Score," representing the similarity between the node and the label (i.e., the node on the other social networking platform), i.e., the similarity between the node in Twitter and the node in LinkedIn. The initial value is 0, ranging from 0 to 1. Through continuous propagation and iteration, the nodes in Twitter and LinkedIn will partially correspond one-to-one, i.e., the maximum similarity score is obtained. Add an "Info" field to store user information. Name the nodes in one knowledge graph as t_*, and the nodes in the other knowledge graph as l_*. For example, the Twitter knowledge graph has the following nodes:

[0031] t_200(Label:l_116,Score:0.5,Info:[name:Bob,birth:2000-12-28……]) represents the Twitter object t_200, whose name is Bob and was born on December 28, 2000. It matches the LinkedIn object l_116 and is tentatively identified as the same person with a similarity of 0.5.

[0032] S1-2, according to the respective social relationship in the respective knowledge graph of the two user account data of Twitter and LinkedIn, an undirected edge is constructed, representing that the two are friends; for example, there is an edge (t_220, t_263) in the Twitter knowledge graph, representing that Twitter users t_220 and t_263 are friends on the Twitter platform.

[0033] S2, false label propagation method based on the degree of nodes in the knowledge graph: the degree of a node in the knowledge graph is defined as the number of edges directly connected to the node; in the Twitter and LinkedIn knowledge graphs, there are a small number of "socially active" nodes, which frequently socialize with most other nodes, so in the knowledge graph, these "socially active" nodes have a large number of edges, so the value of the degree is much larger than that of other nodes. Therefore, a threshold value for the degree is set, and nodes with a degree greater than the threshold value are considered "socially active" nodes; the "socially active" nodes in the two knowledge graphs are arranged in order of the size of the degree, and then the "socially active" nodes in the two knowledge graphs are matched, and the "socially active" nodes in the Twitter knowledge graph are matched with the "socially active" nodes in the LinkedIn knowledge graph whose similarity exceeds the threshold value, and are respectively set as the correct label of the other node (indicating that they belong to the same person's account); and the "socially active" nodes are set with false labels;

[0034] When setting the size of the degree threshold, it should not be too large, otherwise it will result in too few "socially active" nodes being selected, and too few false labels, making it difficult for the label to propagate outward. It should also not be too small, otherwise it may result in ordinary nodes that are not "socially active" being set as "socially active" nodes, and being assigned false labels, resulting in a decrease in label accuracy. Generally, the final threshold setting can filter out 2-3 groups of "socially active" nodes.

[0035] The "socially active" nodes have false labels, which do not mean that they are matched with false labels on the two platforms. The significance of setting false labels is that by setting false labels for these "socially active" nodes, in the subsequent label propagation process, they can use their rich social relationships to facilitate the propagation of labels.

[0036] S3, label propagation method based on user information similarity: taking the "node to be matched" as the center, first locate its "bridge node" (labeled node in the first-order neighbor), then through the matching relationship of the bridge node, use "first-order neighbor set" to construct a set of potential matching nodes, and complete label updating and propagation based on user information similarity.

[0037] In the input data, a small number of nodes already have correct labels (i.e. some active nodes on Twitter and LinkedIn have determined matches with each other), and these nodes spread false labels outward to find more nodes in the two knowledge graphs that do not have labels; the specific implementation steps are as follows:

[0038] S3-1, respectively, sort the nodes in the Twitter and LinkedIn knowledge graphs according to the size of the degree, so that the subsequent steps are in the same order of nodes in the Twitter and LinkedIn knowledge graphs;

[0039] S3-2, according to the size of the degree, traverse the nodes of Twitter (or LinkedIn) from large to small, each traversed node is called a matching node, find out the nodes with correct labels in the first-order neighbors of the node, these nodes are called "bridge nodes" of the matching node, and the "bridge node set" of the matching node is called "bridge node set"; the first-order neighbor is defined as the node directly connected to the matching node;

[0040] S3-3, the "bridge node" of each matching node in the Twitter knowledge graph matches the "bridge node" of the matching node in the LinkedIn knowledge graph, so the potential matching node set of the matching node can be determined by analyzing the "bridge node" of the matching node; the specific method is to find out the nodes in the LinkedIn knowledge graph that match the "bridge nodes" according to the "bridge nodes" of the matching node, and take the union of the first-order neighbors of these nodes as the potential matching node set of the matching node in the LinkedIn knowledge graph, as shown in Figure 2 The bridge nodes of the matching node t_75 in the Twitter knowledge graph include t_486, t_101, etc., and the potential matching nodes of the matching node are l_99, l_145, l_26 in the LinkedIn knowledge graph, l_782 in the LinkedIn knowledge graph is the bridge node of l_99, l_145, l_26, and l_268 is the bridge node of l_99, l_26. The reason for taking the union set is that the data size in the Twitter and LinkedIn knowledge graphs is often inconsistent, and some nodes and edges may be missing in the LinkedIn knowledge graph, which is relatively sparse. If the intersection set is taken, some actual matching nodes may be filtered out because of the lack of edges with a certain bridge node, and cannot enter the potential matching node set, so the union set is taken.

[0041] S3-4, calculate the similarity between the two nodes by the user basic information stored in the "Info" field of the nodes in the set of potential matching nodes, record the node with the highest similarity, called the "most similar node" of the node to be matched;

[0042] If the highest similarity recorded is greater than the existing "Score" label value of the "most similar node", update the existing "Label" label of the "most similar node" to point to the node to be matched, and update the "Score" value to the highest similarity, and reset the "Label" label of the corresponding node in the previous Twitter knowledge graph pointing to the current "most similar node" to "null", and set the "Score" to 0, indicating that it is no longer matched with the "most similar node" of the current node to be matched;

[0043] If the similarity between the current node to be matched and its "most similar node" is less than the existing "Score" label value of the "most similar node", update the "most similar node" of the node to be matched to point to the node in the set of potential matching nodes with the second largest similarity to the node to be matched, and repeat the comparison process described above;

[0044] The calculation of the similarity between the two nodes uses the similarity of the username, user avatar and user profile;

[0045] (1) The similarity of the username is calculated using the Levenshtein edit distance; the Levenshtein edit distance is a method for calculating the similarity between two strings, which measures their similarity by calculating the minimum number of editing operations required to convert one string to another. Let a and b be two usernames, a and b be strings; operation(a, b) means the minimum number of operations required to convert string a to string b, len max (a, b) represents the maximum length of a and b; the similarity S name(a,b) between the two usernames is calculated as follows:

[0046]

[0047] (2) The similarity of the user avatar is calculated using the perceptual hash algorithm; first, all avatars are unified in size and converted to grayscale images, then discrete cosine transform is performed to obtain a DCT coefficient matrix, and a 64-dimensional binary vector is generated as the hash value of the picture; the similarity S of the user avatars is obtained by calculating the Hamming distance between the hash values of the two user avatars pic , the specific calculation formula is as follows:

[0048]

[0049] User personal profile similarity calculation: regarding the calculation of personal profile similarity, the calculation of personal profile similarity is carried out from two dimensions of semantic understanding and text coincidence. First, at the semantic level, the Sentence-Bert model is used to fuse the word vectors in the sentence to directly generate a fixed-length sentence vector, and the semantic similarity S meaning between personal profiles is evaluated through the cosine similarity before the sentence vector; at the text coincidence level, the Jaccard is used to calculate the coincidence degree S Jaccard of the text contents of the two; the semantic similarity and the text coincidence degree are arithmetically averaged to obtain the personal profile similarity:

[0050]

[0051] The username, user avatar and personal profile similarity are arithmetically averaged to obtain the final user information similarity:

[0052]

[0053] S3-5, after each iteration of the above steps, a new matching pair will appear, and the number of matching pairs generated in each round is recorded. When the number of matching pairs is 0, it means that no new node pair can be matched, and the loop is exited, the iteration is stopped, and the label propagation is ended; otherwise, return to step S3-2 to continue matching.

[0054] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present application and should not be understood to limit the scope of the present application to such specific embodiments and examples. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A cross-social network account matching method based on knowledge graph tag propagation, characterized in that, Includes the following steps: S1. Obtain user account data from two social networking platforms respectively, and construct user knowledge graphs for each user account data; the specific steps are as follows: S1-1. Treat each object in the two social networking platforms as a node, add a label "Label" to each node. The label is a node on the other social networking platform, with an initial value of None, and is propagated and iterated through subsequent steps. The label value is "Score", which represents the similarity between the node and the label, with an initial value of 0 and a range of 0 to 1. Add an "Info" field to store user information; name nodes in one knowledge graph as t_* and nodes in another knowledge graph as l_*; S1-2. Construct undirected edges in the knowledge graphs of the two user accounts according to their respective social relationships, representing that the two are friends. S2. A method for propagating false labels based on the degree of nodes in a knowledge graph: The degree of a node in a knowledge graph is defined as the number of edges directly connected to that node. Set a threshold for degree, and designate nodes with a degree greater than the threshold as "socially active" nodes; arrange the "socially active" nodes in the two knowledge graphs in order of degree, and then match the "socially active" nodes in the two knowledge graphs respectively. Match nodes in the two knowledge graphs whose similarity exceeds the threshold and use them as the correct labels for each other's nodes. And set fake labels for these matching "socially active" nodes; S3. Tag propagation method based on user information similarity: In the input data, a small number of nodes already have correct labels. These nodes propagate outwards through false labels, allowing more unlabeled nodes in the two knowledge graphs to find their matching nodes. The specific implementation steps are as follows: S3-1. Sort the nodes in the two knowledge graphs according to their degree. S3-2. In order of degree, traverse any node in the knowledge graph from largest to smallest. Each traversed node is called the node to be matched. Find the nodes with the correct labels among the first-order neighbors of the node to be matched. These nodes are called the "bridging nodes" of the node to be matched. The set of "bridging nodes" is called the "bridging node set" of the node to be matched. The first-order neighbor is defined as the node directly connected to the node to be matched. S3-3. By analyzing the "bridging nodes" of the node to be matched, determine the potential matching node set of the node to be matched. Specifically, based on the "bridging nodes" of the node to be matched, find the nodes in another knowledge graph that match these "bridging nodes", and take the union of the first-order neighbors of these nodes as the potential matching node set of the node to be matched in the other knowledge graph. S3-4. Calculate the similarity between two nodes using the user's basic information stored in the "Info" field of the node to be matched and the nodes in the potential matching node set. Record the node with the highest similarity, called the "most similar node" of the node to be matched. If the recorded similarity is greater than the existing "Score" label value of this "most similar node", update the existing "Label" label of this "most similar node" to point to the node to be matched, and update the "Score" value to the highest similarity. Also, reset the "Label" label of the corresponding node that previously pointed to the current "most similar node" to "null", and set the "Score" to 0. If the similarity between the current node to be matched and its "most similar node" is less than the existing "Score" label value of the "most similar node", then update the "most similar node" of the node to be matched, point it to the node with the second largest similarity to the node to be matched in the potential matching node set, and repeat the above comparison process. S3-5. After each iteration, a new matching pair will appear. Record the number of matching pairs generated in each round. When the number of matching pairs is 0, it means that no new node pair can be matched. Exit the loop, stop the iteration, and end the label propagation.

2. The cross-social network account matching method based on knowledge graph tag propagation according to claim 1, characterized in that, In steps S3-4, the calculation of the similarity of user information between two nodes is achieved by fusing the similarity of username, user avatar and user profile. Username similarity is calculated using Levenshtein edit distance; let the two usernames be a and b, where a and b are both strings. operation(a,b) refers to the minimum number of operations required to convert string a to string b, len max (a, b) represents the maximum length of a and b; then the similarity S between the two usernames is... name(a,b) The calculation method is as follows: User avatar similarity is determined using a perceptual hashing algorithm. First, all avatars are resized and converted to grayscale. Then, a discrete cosine transform (DCT) is performed to obtain the DCT transform coefficient matrix, which is then used to generate a 64-dimensional binary vector. This is used as the hash value of the image; the similarity S between the hash values ​​of two user avatars is obtained by calculating the Hamming distance between them. pic The specific calculation formula is as follows: User profile similarity calculation: At the semantic level, the Sentence-Bert model is used to fuse word vectors in sentences, directly generating fixed-length sentence vectors. The semantic similarity S between profiles is then evaluated using the cosine similarity between the sentence vectors. meaning At the text overlap level, Jaccard is used to calculate the degree of overlap S between the two text contents. Jaccard The semantic similarity and text overlap are arithmetically averaged to obtain the personal profile similarity score. The final user information similarity score is obtained by taking the arithmetic mean of the similarity scores of usernames, profile pictures, and personal profiles.

Images