A method for processing a social network graph

Abstract

The present invention provides a method for processing a social network graph, including: obtaining a network graph to be processed including a first social network graph and a second social network graph, and an initial set of anchor point pairs corresponding to the two graphs; respectively extracting node structure information of the two graphs for the first social network graph and the second social network graph; according to the node structure information of the two graphs, using a heuristic filtering rule to screen out initial anchor point pairs that meet the preset similarity requirements from the initial set of anchor point pairs based on multiple similarity calculation methods, where at least part of the similarity calculation methods calculate the similarity based on relevant features that can represent the neighbor structure in the two graphs; determining the association relationship between nodes in the first social network graph and the second social network graph based on the node structure information and the filtered initial set of anchor point pairs.

Description

Technical Field

[0001] The present invention relates to a method in the field of data mining technology, specifically to the field of hardware acceleration for neural network model calculation, and more specifically, to a method for processing a social network graph. Background Art

[0002] With the development of Internet technology, various social platforms have emerged. The same person (user) often registers on multiple different social platforms and establishes corresponding social networks, thus generating diverse heterogeneous social networks.

[0003] In order to better mine the potential information in these heterogeneous social networks, the corresponding relationships of nodes are usually searched in different social networks, generally also called network alignment, to discover different accounts belonging to the same natural person in multiple social networks, so as to mine potential social relationships; thereby realizing various cross-network applications, such as friend recommendation and content dissemination (such as advertisement / news push, etc.).

[0004] Network alignment needs to utilize some anchor pairs (or called aligned node pairs, that is, node pairs that have been associated in two social networks) to perform supervised network alignment. Existing supervised network alignment methods generally can be divided into two steps. The first step is to use a small known part of aligned node pairs (called anchor pairs) to establish a mapping between networks. Through network representation learning methods such as spectral methods or neural networks, the structural and attribute information of different networks, etc., are encoded into node vectors in the same embedding space. The second step is to align the node vectors in the mapping space according to the similarity function and alignment strategy, and regard the node pair with the highest similarity as the alignment result.

[0005] However, in real scenarios, the network structure differences between social networks and the noise in known anchor pairs bring huge challenges to the alignment task. The network structure differences make the same node may have different performances in different networks, resulting in the emergence of network structure noise, which destroys the preset structural consistency assumption of network alignment; the anchor pair noise affects the mapping between the two networks, making it difficult to accurately judge whether the nodes closest in distance in the embedding space are the same nodes, and the correctness of network alignment needs to be improved. Summary of the Invention

[0006] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for processing a social network graph.

[0007] The purpose of the present invention is achieved by the following technical solutions:

[0008] According to a first aspect of the present invention, there is provided a method for processing a social network graph, including: obtaining two graphs to be processed, namely a first social network graph and a second social network graph, and a set of anchor pairs corresponding to the data of the two graphs; according to the two graphs, using a heuristic filtering rule based on multiple similarity calculation methods to screen out anchor pairs that meet the preset similarity requirements from the set of anchor pairs, obtaining a filtered set of anchor pairs, and this filtered set of anchor pairs is used to determine the correspondence between nodes in the first social network graph and nodes in the second social network graph.

[0009] According to an embodiment of the present invention, the heuristic filtering rule includes: respectively determining a reference value representing similarity for each pair of nodes between the two graphs according to each preset similarity calculation method, and any pair of nodes is an anchor pair or a non-anchor pair; for each pair of nodes, performing corresponding similarity ranking using the reference values obtained by each preset similarity calculation method, obtaining multiple types of similarity rankings; screening according to the performance of each anchor pair in the set of anchor pairs in multiple types of similarity rankings, obtaining a filtered set of anchor pairs.

[0010] According to an embodiment of the present invention, each time of screening, the anchor pairs whose rankings in the multiple types of similarity rankings all meet the preset ranking requirements are used as the anchor pairs that meet the preset similarity requirements.

[0011] According to an embodiment of the present invention, each pair of nodes includes a first node belonging to the first social network graph and a second node belonging to the second social network graph, and the multiple preset similarity calculation methods include the following methods or combinations of the following methods:

[0012] Method 1: Using the PageRank algorithm, determining a first PageRank value of the first node in the first social network graph and a second PageRank value of the second node in the second social network graph, and determining the similarity between the first PageRank value and the second PageRank value as reference value 1;

[0013] Method 2: Obtaining the degree D of the first node in the first social network graph s and the degree D of the second node in the second social network graph t , and taking as reference value 2;

[0014] Method 3: Obtaining the degree D of the first node in the first social network graph s and the degree D of the second node in the second social network graph t , and taking as reference value 3;

[0015] Method 4: Obtain the first clustering coefficient of the first node in the first social network graph and the second clustering coefficient of the second node in the second social network graph, and determine the similarity between the first clustering coefficient and the second clustering coefficient as reference value 4;

[0016] Method 5: Obtain the first neighbor degree vector of the first node in the first social network graph and the second neighbor degree vector of the second node in the second social network, and determine the similarity between the first neighbor degree vector and the second neighbor degree vector as reference value 5;

[0017] Method 6: Obtain the number of occurrences of the first node at different numbered positions in the heterogeneous sub-graph of the first social network graph as the first primitive frequency vector, and the number of occurrences of the second node at different numbered positions in the heterogeneous sub-graph of the second social network graph as the second primitive frequency vector, and determine the similarity between the first primitive frequency vector and the second primitive frequency vector as reference value 6;

[0018] Method 7: Obtain the set of neighbors within the preset hop count range of the first node mapped to the second social network graph according to the current set of anchor point pairs to form the mapped neighbor set of the first node in the second social network graph, and the neighbor set of the second node in the second social network graph, and determine the similarity between the mapped neighbor set and the neighbor set as reference value 7.

[0019] According to an embodiment of the present invention, any neighbor degree vector in the first neighbor degree vector and the second neighbor degree vector is obtained by multiplying the number of neighbor nodes satisfying the preset condition at each hop count within the preset hop count range in the corresponding social network graph by the distance attenuation coefficient preset for each hop count, where the preset condition is that the value obtained by taking the logarithm of the degree of the corresponding neighbor node at this hop count is within the preset screening range value.

[0020] According to an embodiment of the present invention, reference value 6 is determined in the following manner:

[0021] Reference value

[0022] where, gl s represents the first primitive frequency vector, gl t represents the second primitive frequency vector, max(gl s , gl t ) represents taking the maximum value from gl s and gl s .

[0023] According to a second aspect of the present invention, there is provided a method for aligning nodes of a social network graph, including: S1. Perform an initial alignment operation, including: obtaining two graphs to be processed and an initial set of anchor pairs, and obtaining a filtered initial set of anchor pairs therefrom by using the method described in the first aspect; respectively performing graph embedding processing on the first social network graph and the second social network graph by using the filtered initial set of anchor pairs and a random walk model with restart, obtaining the feature vectors of the nodes in each graph, and performing alignment based on the distances between the feature vectors of each pair of nodes in the two graphs to obtain an alignment result; S2. Perform one or more intermediate alignment operations, including: using the latest alignment result as an intermediate set of anchor pairs, and obtaining a filtered intermediate set of anchor pairs therefrom by using the method described in the first aspect; respectively performing graph embedding processing on the first social network graph and the second social network graph by using the filtered intermediate set of anchor pairs and a random walk model with restart, updating the feature vectors of the nodes in each graph, and performing alignment based on the distances between the updated feature vectors of each pair of nodes in the two graphs to obtain an updated alignment result.

[0024] According to a second aspect of the present invention, there is provided a method for aligning nodes of a social network graph, including: S1. Perform an initial alignment operation, including: obtaining two graphs to be processed and an initial set of anchor pairs, and obtaining a filtered initial set of anchor pairs therefrom by using the method described in the first aspect; respectively performing graph embedding processing on the first social network graph and the second social network graph by using the filtered initial set of anchor pairs and a random walk model with restart, obtaining the feature vectors of the nodes in each graph, and performing alignment based on the distances between the feature vectors of each pair of nodes in the two graphs to obtain an alignment result; S2. Perform one or more intermediate alignment operations, including: using the latest alignment result as an intermediate set of anchor pairs, and obtaining a filtered intermediate set of anchor pairs therefrom by using the method described in the first aspect; respectively performing graph embedding processing on the first social network graph and the second social network graph by using the filtered intermediate set of anchor pairs and a random walk model with restart, updating the feature vectors of the nodes in each graph, and performing alignment based on the distances between the updated feature vectors of each pair of nodes in the two graphs to obtain an updated alignment result; S3. Perform a minimum weight matching of a bipartite graph on the difference set between the last alignment result and its filtered alignment result to delete duplicate corresponding relationships, obtaining a matching result, where the filtered alignment result is obtained by processing the last alignment result by using a heuristic filtering rule; S4. Use the union of the matching result and the last obtained filtered alignment result as the final alignment result.

[0025] According to a fourth aspect of the present invention, there is provided an electronic device, including: one or more processors; and a memory, where the memory is used to store executable instructions; the one or more processors are configured to implement the steps of the method described in the first aspect by executing the executable instructions.

[0026] Compared with the prior art, the advantages of the present invention are as follows:

[0027] The present invention adopts heuristic filtering rules to screen out initial anchor pairs that meet the preset similarity requirements from the set of initial anchor pairs based on multiple similarity calculation methods, and obtains a filtered set of initial anchor pairs. By using the filtered set of initial anchor pairs, the corresponding relationship between at least some nodes in the first social network graph and the second social network graph is determined. Thus, a part of the anchor pair noise in the set of initial anchor pairs can be eliminated according to multiple similarity calculation methods, so as to eliminate the influence of a part of the network structure noise during alignment and improve the alignment accuracy of the social network graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The following further describes embodiments of the present invention with reference to the accompanying drawings, where:

[0029] Figure 1 FIG. is a flowchart of a method for aligning nodes of a social network graph according to an embodiment of the present invention;

[0030] Figure 2 FIG. is a flowchart of a method for processing a social network graph according to an embodiment of the present invention;

[0031] Figure 3 FIG. is a schematic diagram for explaining a subgraph and a graph element frequency vector according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below through specific embodiments with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0033] As mentioned in the background art section, the existence of network structure noise and anchor pair noise leads to the need to improve the correctness of network alignment. When conducting research on network structure (i.e., the structure of the social network graph) alignment, the inventor repeatedly studied and determined that the problem of the prior art is caused by the difficulty in distinguishing noise from true alignment node pairs during the alignment process. In this regard, through further research on network structure noise and anchor pair noise, the inventor found that preferentially filtering out high-confidence anchor pairs during the alignment process is beneficial to correcting the network structure noise, that is, reducing anchor pair noise helps to correct the network structure noise, so as to further improve the alignment accuracy rate during alignment. In this regard, the present invention adopts a heuristic filtering rule to screen out initial anchor pairs that meet the preset similarity requirements from the initial anchor pair set based on multiple similarity calculation methods, obtaining a filtered initial anchor pair set, and determining the corresponding relationship between at least some nodes in the first social network graph and the second social network graph by using the filtered initial anchor pair set. Thus, a part of the anchor pair noise in the initial anchor pair set can be eliminated based on multiple similarity calculation methods, so as to eliminate the influence of a part of the network structure noise during alignment and improve the alignment accuracy rate of the social network graph alignment.

[0034] According to an embodiment of the present invention, referring to Figure 1 , a method for aligning nodes of a social network graph is provided, including steps A1, A2, A3, and A4, and each step will be described in detail below with specific embodiments.

[0035] Step A1: Perform an initial alignment operation. The initial alignment operation includes: obtaining two graphs to be processed and an initial anchor pair set, and obtaining a filtered initial anchor pair set therefrom by using the processing method of the social network graph; performing graph embedding processing on the first social network graph and the second social network graph respectively by using the filtered initial anchor pair set and the random walk model with restart, obtaining the feature vectors of the nodes in each graph and performing alignment based on the distances between the feature vectors of every two nodes in the two graphs, and obtaining an alignment result.

[0036] According to an embodiment of the present invention, schematically, the processing method of the social network graph includes step A111 and step A112, which will be described separately below:

[0037] Step A111: Obtain two graphs (Graphs) to be processed, namely the first social network graph and the second social network graph, and the initial anchor pair set corresponding to the data of the two graphs;

[0038] According to an embodiment of the present invention, a social network graph is a social network represented by a graph data structure (Graph). In the graph data structure, nodes represent users and edges define the relationships between users. The first social network graph and the second social network graph can be directed graphs or undirected graphs, and the embodiments of the present invention do not make specific limitations here. The first social network graph and the second social network graph can be social network graphs obtained from different social platforms respectively. For example, the first social network graph can be a social network graph formed among users on the Sina Weibo platform, while the second social network graph can be a social network graph formed among users on the QQ platform. Of course, it can also be other types of social platforms, and the embodiments of the present invention do not make specific limitations here.

[0039] According to an embodiment of the present invention, an anchor pair set refers to a pair of nodes that have a corresponding relationship in the first social network graph and the second social network graph. In this application, the anchor pair set is divided into an initial anchor pair set and an intermediate anchor pair set. Among them, the initial anchor pair set is the anchor pair pre-labeled for the first social network graph and the second social network graph. The intermediate anchor pair set is a pair of nodes that have a corresponding relationship in the first social network graph and the second social network graph after node alignment, that is, the alignment result, which will be described later and will not be elaborated here.

[0040] Step A112: According to the two graphs, use heuristic filtering rules to screen out anchor pairs that meet the preset similarity requirements from the initial anchor pair set based on multiple similarity calculation methods, and obtain a filtered initial anchor pair set.

[0041] According to an embodiment of the present invention, in step A112, the initial anchor pair set is filtered using heuristic filtering rules to obtain a filtered initial anchor pair set.

[0042] See Figure 2, for any set of anchor point pairs, the processing process of the heuristic filtering rule includes: Q1. Obtain two graphs to be processed, namely the first social network graph and the second social network graph, and the set of anchor point pairs corresponding to the data of the two graphs; Q2. According to the two graphs, use the heuristic filtering rule to screen out the anchor point pairs that meet the preset similarity requirements from the set of anchor point pairs based on multiple similarity calculation methods, and obtain the filtered set of anchor point pairs. In step A112, use the heuristic filtering rule to determine a reference value representing similarity for each pair of nodes formed by every two nodes between the two graphs according to each preset similarity calculation method. Any pair of nodes is an anchor point pair or a non-anchor point pair; for each pair of nodes, perform corresponding similarity ranking using the reference values obtained by each preset similarity calculation method to obtain multiple types of similarity rankings; screen according to the performance of each anchor point pair in the initial set of anchor point pairs in multiple types of similarity rankings to obtain the filtered initial set of anchor point pairs. It should be understood that there are various ways to evaluate the performance in the similarity ranking. Preferably, each time of screening, the anchor point pairs whose rankings in the multiple types of similarity rankings all meet the preset ranking requirements are used as the anchor point pairs that meet the preset similarity requirements. For example: set a ranking threshold, which is a value obtained by multiplying the total ranking by a predetermined ratio (such as 60%, 65%, 70%, 75% or 80%, etc.). If the similarity ranking is obtained from high to low according to similarity, then when the ranking of the corresponding anchor point pair in a type of similarity ranking is less than or equal to the ranking threshold, the ranking of the anchor point pair in this type of similarity ranking meets the preset ranking requirements; if the rankings of the anchor point pair in all types of similarity rankings all meet the corresponding preset ranking requirements, then the anchor point pair is an anchor point pair that meets the preset similarity requirements. It should be understood that in addition, those skilled in the art can also use other methods for evaluation. For example, set weights for each type of similarity ranking, and perform weighting on the multiple types of similarity rankings based on the weights to obtain a weighted similarity ranking, and regard the anchor point pairs with the weighted similarity ranking less than or equal to the ranking threshold as the anchor point pairs that meet the preset similarity requirements.

[0043] According to an embodiment of the present invention, each pair of nodes includes a first node belonging to the first social network graph and a second node belonging to the second social network graph. The multiple preset similarity calculation methods include the following methods or combinations of the following methods (i.e., combinations of Method 1 - Method 7):

[0044] Method 1: Using the PageRank algorithm, determine the first PageRank value of the first node in the first social network graph and the second PageRank value of the second node in the second social network graph, and determine the similarity between the first PageRank value and the second PageRank value as reference value 1. It should be understood that the PageRank algorithm was originally used to calculate the web page ranking (i.e., the PageRank value of the web page). However, the nodes of the graph data can be regarded as web pages, and the connections between nodes (i.e., edges) can be regarded as the links of web pages (i.e., jump links between web pages), then the PageRank algorithm can be used to determine the PageRank value of the nodes in the social network graph. In addition, there are also multiple calculation versions of the PageRank algorithm, which can be used according to the needs of the implementer, and this embodiment does not make any limitations in this regard. Through Method 1, it is helpful to screen high-confidence anchor pairs from the perspective of the similarity of the PageRank values of the nodes to reduce the noise of the anchor pairs.

[0045] Method 2: Obtain the degree D of the first node in the first social network graph s and the degree D of the second node in the second social network graph t , and use as reference value 2. The degree of a node in a graph (Degree), abbreviated as the degree of the node, also known as the correlation degree, is a basic concept of the graph structure, which refers to the number of edges associated with the node. Through Method 2, it is helpful to screen high-confidence anchor pairs from the perspective of the relative similarity of the degrees of the nodes to reduce the noise of the anchor pairs.

[0046] Method 3: Obtain the degree D of the first node in the first social network graph s and the degree D of the second node in the second social network graph t , and use as reference value 3. Through Method 3, it is helpful to screen high-confidence anchor pairs from the perspective of the absolute similarity of the degrees of the nodes to reduce the noise of the anchor pairs.

[0047] Method 4: Obtain the first clustering coefficient of the first node in the first social network graph and the second clustering coefficient of the second node in the second social network graph, and determine the similarity between the first clustering coefficient and the second clustering coefficient as reference value 4. In graph theory, the clustering coefficient of a node (Clustering coefficient) refers to the degree of clustering of the neighbors of the node in the graph (i.e., the local clustering coefficient). Through Method 4, it is helpful to screen high-confidence anchor pairs from the perspective of the similarity of the local clustering coefficients to reduce the noise of the anchor pairs.

[0048] Method 5: Obtain the first neighbor degree vector of the first node in the first social network graph and the second neighbor degree vector of the second node in the second social network, and determine the similarity between the first neighbor degree vector and the second neighbor degree vector as reference value 5. Through Method 5, it helps to screen high-confidence anchor pairs from the perspective of the similarity of neighbor degree vectors to reduce anchor pair noise. Preferably, any neighbor degree vector in the first neighbor degree vector and the second neighbor degree vector is obtained by multiplying the number of neighbor nodes satisfying the preset condition at each hop within the preset hop range in the corresponding social network graph by the preset distance attenuation coefficient for each hop. Wherein, the preset condition is that the value obtained by taking the logarithm of the degree of the corresponding neighbor node at this hop is within the preset screening range value. For example, taking node s as an example, for all neighbor nodes at a distance of i hops from node s, after calculating the log2D value of their degrees D and rounding up to an integer value as the degree information of this neighbor node, taking the number of neighbor nodes at i hops with degree information from 0 to m as the degree vector of the i-hop neighbor; for nodes within 3 hops from s, 3 different degree vectors are obtained. For the degree vector of k hops, multiply by 0.5 k-1 as the distance attenuation coefficient, and then accumulate the weighted degree vectors as the value of the neighbor degree vector of node s. The similarity of neighbor degree vectors can use existing algorithms. For example, use the Radial Basis Function (RBF) kernel to calculate the similarity between the neighbor degree vectors of the two nodes in the node pair (that is, the first neighbor degree vector and the second neighbor degree vector) to obtain reference value 5. Using the distance attenuation coefficient can pay more attention to the influence of the values of the degree vectors of neighbors closer to the node, and improve the accuracy of filtering based on reference value 5. It should be understood that it is also feasible in principle not to set the distance attenuation coefficient to generate other implementation manners.

[0049] Method 6: Obtain the number of times the first node appears at different numbered positions in the heterogeneous subgraph of the first social network graph as the first primitive frequency vector, and the number of times the second node appears at different numbered positions in the heterogeneous subgraph of the second social network graph as the second primitive frequency vector, and determine the similarity between the first primitive frequency vector and the second primitive frequency vector as reference value 6. The topological structure of each numbered node on the primitive is unique. Count the number of times the first node s and the second node t appear at different numbered positions in the primitive, and convert it into the primitive frequency vectors gl s and gl t . For easy understanding, see Figure 3 , and illustrate the subgraph and the primitive frequency vector through a simplified graph structure. Assume a graph is composed of four nodes a, b, c, and d, and the connection relationship is as Figure 3As shown in the graph structure. Correspondingly, there are 3 types of heterogeneous subgraphs in this graph structure, namely subgraph x, subgraph y, and subgraph z. The following shows the distribution of graph elements with the number of nodes a and b within 3: The top row shows non-isomorphic subgraphs with the number of nodes within 3. Gray and white are used to distinguish nodes in different positions, and nodes with the same topological structure are represented by the same color; below each subgraph is the distribution of graph elements of nodes a and b, and the subgraphs are separated by gray dashed lines; in each graph element schematic diagram, the bold circles highlight nodes a and b, the solid lines represent the structures related to the current nodes and graph elements, and the dashed lines represent unrelated structures. The numbers next to the nodes represent the graph element numbers. For graph element 0, node a appears 3 times in 3 structures: (a,b), (a,d), and (a,c). Therefore, the 0th dimension of the graph element frequency vector of a is 3; for graph element 2, node b appears 1 time in the structure (a,b,d). Therefore, the 2nd dimension of the graph element frequency vector of b is 1. According to this rule, for nodes a and b, the graph element degree vectors with the number of nodes within 3 are (3, 2, 3, 1) and (2, 2, 1, 1) respectively, as shown in Figure 3 the table on the left. The situation of nodes c and d is similar and will not be elaborated here. Preferably, the reference value 6 is determined in the following manner:

[0050] Reference value

[0051] where gl s represents the first graph element frequency vector, gl t represents the second graph element frequency vector, and max(gl s , gl t ) represents taking the maximum value from gl s and gl s . Through method 5, it helps to screen high-confidence anchor pairs from the perspective of the similarity of graph element frequency vectors to reduce anchor pair noise.

[0052] Method 7: Obtain the mapping neighbor set of the first node within the preset hop count range according to the current set of anchor pairs to the second social network graph to form the mapping neighbor set of the first node in the second social network graph, and the neighbor set of the second node in the second social network graph, and determine the similarity between the mapping neighbor set and the neighbor set as reference value 7. The neighbor set can be represented by the numbers of the neighbor nodes of the node. Through method 7, it helps to screen high-confidence anchor pairs from the perspective of the similarity of the neighbor sets of nodes to reduce anchor pair noise. It should be understood that inspired by the present invention, implementers can also add, delete, or adjust some similarity calculation methods to obtain other similar implementation manners.

[0053] According to an embodiment of the present invention, the filtered initial anchor point pair set and the random walk with restart model are respectively used to perform graph embedding processing on the first social network graph and the second social network graph, obtain the feature vectors of the nodes in each graph, and align them based on the distances between the feature vectors of every two nodes in the two graphs to obtain the alignment result. The steps are illustrated by the following schematic embodiments A121 and A122.

[0054] A121: Use the filtered initial anchor point pair set and the random walk with restart model to perform graph embedding processing on the first social network graph and the second social network graph respectively, and obtain the feature vectors of the nodes in each graph.

[0055] According to an embodiment of the present invention, the step of performing graph embedding processing on a social network graph using a random walk model can be adopted. However, in order to improve the anti-interference ability, further reduce the influence of structural noise, and improve the alignment accuracy rate, the random walk with restart model is adopted in this embodiment to perform graph embedding processing on the social network graph. Schematically, the schematic process of performing embedding processing on any graph to obtain the feature vectors of the nodes in the graph includes: inputting the adjacency matrix A of the graph G to be aligned; after performing (or executing) row L1 normalization on A, obtaining the transition probability matrix W; for the input restart probability c, the random walk with restart probability matrix R = (1 - c)(I - cW) -1 , where I is the identity matrix; each row of R is the transition probability vector from a node in graph G to all other nodes (which can also be called the random walk with restart vector and can be used to represent the structural information of the node in graph G); when performing node alignment on networks G s and G t , use the anchor point pair set A ′ = {(s, t)|s ∈ S, t ∈ T} to establish a mapping (i.e., a corresponding relationship) between two different graphs, where S and T are the anchor node sets of graphs G s and G t respectively; for the nodes in G s , use the transition probability vector from the node to all nodes in the anchor node set S as the feature vector of the node; perform the same operation on the nodes in G t to obtain the corresponding feature vectors of the nodes, which will not be elaborated here. The technical solution of this embodiment can at least achieve the following beneficial technical effects: The random walk with restart has stronger anti-interference ability compared with general spectral methods, can reduce the influence of structural noise, and has a high alignment accuracy rate.

[0056] A122: Align based on the distances between the feature vectors of every two nodes in the two graphs to obtain the alignment result.

[0057] According to an embodiment of the present invention, for each node in the first social network graph, the node in the second social network graph with the smallest distance from the feature vector is used as the corresponding node of this node in the first social network graph to obtain an alignment result. Schematically, when aligning, for example, the Manhattan distance (Cityblock Distance) is used to measure the distance between two feature vectors as the similarity of the feature vectors of two nodes; for each node in G s the node in G t with the highest similarity (i.e., the smallest Manhattan distance) to it is used as its alignment result.

[0058] Step A2: Perform one or more intermediate alignment operations. Each intermediate alignment operation includes: using the latest alignment result as the set of intermediate anchor pairs, and obtaining a filtered set of intermediate anchor pairs therefrom by using the processing method of the social network graph; using the filtered set of intermediate anchor pairs and the random walk model with restart to perform graph embedding processing on the first social network graph and the second social network graph respectively, updating the feature vectors of the nodes in each graph, and aligning based on the distances between the updated feature vectors of every two nodes in the two graphs to obtain an updated alignment result.

[0059] According to an embodiment of the present invention, the difference between the intermediate alignment operation and the initial alignment operation is that the initial set of anchor pairs in the initial alignment operation is replaced by the set of intermediate anchor pairs, so it will not be elaborated here. The number of intermediate alignment operations can be set by the implementer according to needs, such as setting it to a certain value, or adopting a predetermined evaluation index (such as the difference between multiple alignment results is less than a predetermined difference threshold) to determine whether to stop the intermediate alignment operation. The technical solution of this embodiment can at least achieve the following beneficial technical effects: After one or more intermediate alignment operations, higher-confidence anchor pairs can be further filtered out, which is beneficial to correcting the structural noise of the network and improving the accuracy of the alignment result.

[0060] Step A3: Perform a bipartite graph minimum weight matching on the difference set between the last alignment result and its filtered alignment result to delete duplicate corresponding relationships to obtain a matching result, where the filtered alignment result is obtained by processing the last alignment result using a heuristic filtering rule.

[0061] According to an embodiment of the present invention, the result confidence in the difference set is relatively low. Therefore, after multiple rounds of alignment and filtering, re-alignment is performed on a part of the nodes with low confidence in the alignment result. For this, the minimum weight matching algorithm of the bipartite graph can be used (for example, the method in the Karp paper, see Richard Manning Karp: An algorithm to Solve the m x n Assignment Problem in Expected Time O(mn log n). Networks, 10(2): 143–152, 1980.) to delete duplicate correspondence relationships to improve the accuracy of the final alignment result. Schematically, input the set of low-confidence node pairs (i.e., the said difference set) L = {(u, v)|u ∈ U, v ∈ V}, and construct the bipartite graph G = {U, V, E}, where U and V are the node sets of two different graphs respectively, u and v represent the nodes therein, and E is the edge of the weighted bipartite graph, which is constructed by the following method: for each node u in the node set U, select the top k nodes with the highest similarity, add edges between u and these nodes, and add the distance of the feature vectors between the nodes as the weight of the edge to E. The same operation is taken for each node v in V. After constructing the bipartite graph, perform the minimum weight matching algorithm of the bipartite graph on the bipartite graph G, and delete the edges on the bipartite graph G to minimize the sum of the weights of all edges to obtain the matching result. The technical solution of this embodiment can at least achieve the following beneficial technical effects: during the alignment process, the alignment between high-confidence node pairs is relatively reliable, while for low-confidence node pairs, especially the alignment of duplicate nodes is very likely to be incorrect. Therefore, the minimum weight matching method of the bipartite graph with one-to-one alignment is specifically used for processing to ensure that there will be no duplicate alignment results. Since the number of duplicate nodes is reduced, the alignment accuracy rate will further increase.

[0062] Step A4: Use the union of the matching result and the filtered alignment result corresponding to the last alignment result as the final alignment result.

[0063] According to an embodiment of the present invention, that is, append the matching result to the filtered alignment result corresponding to the last alignment result and output the final alignment result.

[0064] For verification, the applicant conducted experiments on the above embodiment. In the Arenas email network alignment task, after the processing of steps A1 - A4, for the case of up to 20% structural differences and 40% anchor node noise, the accuracy rate can reach 98.59%.

[0065] In addition, it should be understood that, except for the foregoing embodiments, those skilled in the art can also use other methods to determine the correspondence between the nodes in the first social network graph and the nodes in the second social network graph by using the filtered set of anchor pairs.

[0066] According to another embodiment of the present invention, there is provided a method for aligning nodes of a social network graph, including steps: A1, A2, A3', where steps A1 and A2 are the same as steps A1 and A2 of the foregoing embodiments and will not be elaborated herein. Step A3' includes: taking the last alignment result as the final alignment result. That is: the minimum weight matching algorithm of the bipartite graph is not used to remove duplicates. However, in the case of performing an initial alignment operation and one or more intermediate alignment operations based on a heuristic filtering rule, the accuracy of the alignment result can still be improved.

[0067] Or, according to still another embodiment of the present invention, there is provided a method for aligning nodes of a social network graph, including steps: A1', A2, A3', where the difference between step A1' and A1 of the foregoing embodiment is that a heuristic filtering rule is not used to filter the initial set of anchor pairs, that is, an initial alignment operation is directly performed based on the initial set of anchor pairs. Only in step A2, a heuristic filtering rule is used to filter the intermediate set of anchor pairs.

[0068] It should be noted that although the above steps are described in a specific order, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even the order can be changed as long as the required functions can be achieved.

[0069] The present invention can be a system, a method, and / or a computer program product. The computer program product can include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0070] The computer-readable storage medium can be a tangible device that retains and stores instructions for use by an instruction execution device. The computer-readable storage medium can for example include, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the above.

[0071] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is also not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of technologies in the market, or to enable other ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for processing a social network graph, characterized in that Including: Obtain two graphs to be processed, namely a first social network graph and a second social network graph, as well as a set of anchor pairs corresponding to the data of the two graphs; According to the two graphs, use heuristic filtering rules to screen out anchor pairs that meet the preset similarity requirements from the set of anchor pairs based on multiple similarity calculation methods, and obtain a filtered set of anchor pairs. This filtered set of anchor pairs is used to determine the correspondence between nodes in the first social network graph and nodes in the second social network graph. Among them, the heuristic filtering rules include: For each preset similarity calculation method, determine a reference value representing similarity for each node pair formed by every two nodes between the two graphs. Any node pair is either an anchor pair or a non-anchor pair; For each node pair, perform corresponding types of similarity ranking using the reference values obtained by each preset similarity calculation method to obtain multiple types of similarity rankings; and Screen according to the performance of each anchor pair in the set of anchor pairs in multiple types of similarity rankings to obtain a filtered set of anchor pairs.

2. The method according to claim 1, characterized in that, Each time of screening, use the anchor pairs whose rankings in the multiple types of similarity rankings all meet the preset ranking requirements as the anchor pairs that meet the preset similarity requirements; or, each time of screening, weight the multiple types of similarity rankings based on the weights of each type of similarity ranking to obtain a weighted similarity ranking, and regard the anchor pairs with the weighted similarity ranking less than or equal to the ranking threshold as the anchor pairs that meet the preset similarity requirements.

3. The method according to claim 1, wherein Each node pair includes a first node belonging to the first social network graph and a second node belonging to the second social network graph. The multiple preset similarity calculation methods include the following methods or combinations of the following methods: Method 1: Use the PageRank algorithm to determine the first PageRank value of the first node in the first social network graph and the second PageRank value of the second node in the second social network graph, and determine the similarity between the first PageRank value and the second PageRank value as reference value 1; Method 2: Obtain the degree of the first node in the first social network graph and the degree of the second node in the second social network graph , and use as reference value 2; Method 3: Obtain the degree of the first node in the first social network graph and the degree of the second node in the second social network graph , and use as reference value 3; Method 4: Obtain the first clustering coefficient of the first node in the first social network graph and the second clustering coefficient of the second node in the second social network graph, and determine the similarity between the first clustering coefficient and the second clustering coefficient as reference value 4; Method 5: Obtain the first neighbor degree vector of the first node in the first social network graph and the second neighbor degree vector of the second node in the second social network, and determine the similarity between the first neighbor degree vector and the second neighbor degree vector as reference value 5; Method 6: Obtain the number of occurrences of the first node at different numbered positions in the heterogeneous subgraph of the first social network graph as the first graph element frequency vector, and the number of occurrences of the second node at different numbered positions in the heterogeneous subgraph of the second social network graph as the second graph element frequency vector, and determine the similarity between the first graph element frequency vector and the second graph element frequency vector as reference value 6; Method 7: Obtain the neighbor set within the preset hop count range of the first node mapped to the second social network graph according to the current anchor pair set to form the mapped neighbor set of the first node in the second social network graph, and the neighbor set of the second node in the second social network graph, and determine the similarity between the mapped neighbor set and the neighbor set as reference value 7.

4. The method according to claim 3, wherein Any of the first neighbor degree vector and the second neighbor degree vector is obtained by multiplying the number of neighbor nodes satisfying the preset condition at each hop within the preset hop count range in the corresponding social network graph by the distance attenuation coefficient preset for each hop. Among them, the preset condition is that the value obtained by taking the logarithm of the degree of the corresponding neighbor node at this hop is within the preset screening range value.

5. The method according to claim 3, characterized in that Reference value 6 is determined in the following manner: Among them, represents the first primitive frequency vector, represents the second primitive frequency vector, represents taking the maximum value from and to obtain the maximum value.

6. A method for node alignment of a social network graph, characterized in that, Including: S1. Perform an initial alignment operation, including: Obtain two graphs to be processed and an initial anchor pair set, and obtain a filtered initial anchor pair set therefrom by using the method according to any one of claims 1-5; Perform graph embedding processing on the first social network graph and the second social network graph respectively by using the filtered initial anchor pair set and the random walk model with restart, obtain the feature vectors of the nodes in each graph, and perform alignment based on the distances between the feature vectors of every two nodes in the two graphs to obtain an alignment result; S2. Perform one or more intermediate alignment operations, including: Take the latest alignment result as the intermediate anchor pair set, and obtain a filtered intermediate anchor pair set therefrom by using the method according to any one of claims 1-5; Perform graph embedding processing on the first social network graph and the second social network graph respectively by using the filtered intermediate anchor pair set and the random walk model with restart, update the feature vectors of the nodes in each graph, and perform alignment based on the distances between the updated feature vectors of every two nodes in the two graphs to obtain an updated alignment result.

7. A method for node alignment of a social network graph, characterized in that, Including: S1. Perform an initial alignment operation, including: Obtain two graphs to be processed and an initial anchor pair set, and obtain a filtered initial anchor pair set therefrom by using the method according to any one of claims 1-5; Perform graph embedding processing on the first social network graph and the second social network graph respectively by using the filtered initial anchor pair set and the random walk model with restart, obtain the feature vectors of the nodes in each graph, and perform alignment based on the distances between the feature vectors of every two nodes in the two graphs to obtain an alignment result; S2. Perform one or more intermediate alignment operations, including: Take the latest alignment result as the intermediate anchor pair set, and obtain a filtered intermediate anchor pair set therefrom by using the method according to any one of claims 1-5; Perform graph embedding processing on the first social network graph and the second social network graph respectively by using the filtered intermediate anchor pair set and the random walk model with restart, update the feature vectors of the nodes in each graph, and perform alignment based on the distances between the updated feature vectors of every two nodes in the two graphs to obtain an updated alignment result; S3. Perform a minimum weight matching of a bipartite graph on the difference set between the last alignment result and its filtered alignment result to remove duplicate corresponding relationships, obtaining a matching result, where the filtered alignment result is obtained by processing the last alignment result using a heuristic filtering rule; S4. Use the union of the matching result and the last obtained filtered alignment result as the final alignment result.

8. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method according to any one of claims 1 to 7.

9. An electronic device, characterized in that, Comprising: One or more processors; And A memory, where the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method according to any one of claims 1 to 7 by executing the executable instructions.

Images