Fusion multi-social media platform knowledge graph neighborhood structure non-heterogeneous entity alignment method

Abstract

The application relates to a kind of knowledge graph neighborhood structure non-isomorphism entity alignment methods of fusion multi-social media platform, comprising: collecting and fusing the data of multi-social media platform, and constructing knowledge graph according to the interactive behavior and theme content of user;The attention weight between entities is calculated, and the neighbor entity that contributes more to the center entity is obtained;By topological invariant function, the single-hop and two-hop neighbors of entity are aggregated respectively, and the single-hop and two-hop neighbor representation information of entity is supplemented;Using the gate mechanism combines single-hop and two-hop neighbor representation, obtains entity representation;According to the similarity between entities calculated by alignment module, the result of entity alignment is obtained.The application uses neighborhood structure non-isomorphism entity alignment model to weaken the feature loss of non-isomorphism entity in graph structure conversion and traditional graph convolution network conversion, solves the problem caused by entity non-isomorphism, so that the result of entity alignment is more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine learning, and in particular to an entity alignment method for knowledge graph neighborhood structure non-isomorphism fusion of multi-social media platform data. BACKGROUND

[0002] With the increasing popularity of online social media in recent years, the historical user data of users on various media platforms contains a variety of different time, place, person, event, content and other information, which constitutes a knowledge graph represented as a multi-dimensional heterogeneous network. Users may use different social media platforms to distinguish different needs, so the same user may generate different information records on multiple different social media platforms. For example, a user can use Facebook to share recent life anecdotes and moods with friends, and use Twitter to understand the latest news and current affairs. However, the accounts and statements made by the same user on different social media platforms are mostly isolated and have no association with each other.

[0003] The process of determining whether the user entities in the knowledge graphs constructed by different social media platforms point to the same user in the real world is called user entity alignment. Aligning user entities in knowledge graphs constructed by different social media platforms has great value for many cross-platform applications in real life. For example, after aligning the user entities of Facebook and Twitter, the personal interests and social relationships of a user corresponding to the user entity on Facebook can be analyzed, and then the analysis results can be used to push new topics that the user may be interested in or new friends that the user may follow to the user's Twitter.

[0004] Current user entity alignment solutions mainly rely on the graph structure of the knowledge graph and are based on the assumption that the entities of the multi-platform social media data knowledge graph have neighborhood isomorphism characteristics. These methods have achieved first-class performance on some synthetic data sets extracted from large-scale knowledge graphs, but synthetic data sets are generated after extraction and arrangement of real data sets, and current alignment algorithms cannot produce satisfactory results on actual data sets with diverse structures.

[0005] Only a small part of the entities in real-life knowledge graphs are strictly isomorphic and have frequent connections and rich communication, while most entities only contain a small amount of organizational information and have sparse connections with distant neighbors. Therefore, the current entity alignment algorithm only assumes information isomorphism, which does not perform well on real data, and the influence of entity features in the knowledge graph construction process needs to be considered. SUMMARY

[0006] Technical problems to be solved

[0007] In view of the fact that the entity alignment result of the knowledge graph neighborhood structure of the fusion multi-social media platform is not accurate enough, an entity alignment model of neighborhood structure non-isomorphism is proposed, and a graph convolutional neural network with a transformation invariant function and a graph attention mechanism are introduced. This method can maximize the preservation of the original features of the entities in the knowledge graph, accurately represent the information of the nodes, and improve the accuracy of the alignment result.

[0008] Technical scheme

[0009] An entity alignment method for the neighborhood structure non-isomorphism of a knowledge graph of a fusion multi-social media platform, characterized in that the steps are as follows:

[0010] S1. Constructing a knowledge graph according to the interaction and interaction theme content of user entities on multiple social media platforms;

[0011] S2. Calculating the attention weight between entities and marking the neighbor entities that contribute more to the center entity;

[0012] S3. Aggregating the single-hop and two-hop neighbors of the entity through a topological invariant function to supplement the single-hop and two-hop neighbor representation information of the entity;

[0013] S4. Obtaining the entity representation by combining the single-hop and two-hop neighbor representation using a gating mechanism;

[0014] S5. Obtaining the final alignment result by calculating the similarity between entities according to the processed entity representation of the alignment module.

[0015] Further, in the knowledge graph constructed in S1, the user and the comment theme are used as entity nodes, and the user and the user, the user and the comment theme are used in the form of a triple <h, r, t> to form a knowledge graph.

[0016] Further, in S2, the method for calculating the attention weight between entities is as follows:

[0017] S21. Introducing a graph attention network according to the knowledge graph;

[0018] S22. Applying the graph attention network to the center entity and the two-hop and above neighbor nodes;

[0019] S23. Using matrices and to perform linear transformation on the center entity and the neighbor entity respectively to obtain the attention weight of the entity;

[0020] S24. Using a softmax function to normalize the attention weight of the neighbor nodes of the entity.

[0021] Further, in S3, the entity embedding of the knowledge graph through the topological invariant function is as follows:

[0022] S31: Extracting a feature matrix of the entity according to the knowledge graph and the entity representation processed in S2

[0023] S32: Transforming the feature matrix Input transformation invariant function Mask gate MG (l) In the middle;

[0024] S33: Taking the output value MG (l) (H (l) ) of the function in the previous step as input, and inputting into a graph convolutional neural network;

[0025] S34: Applying a transformation invariant function Highway gate to the original features and the linearly transformed features for interpolation;

[0026] S35: Outputting the knowledge graph after the entity embedding based on the transformation invariant function of the graph convolutional neural network.

[0027] Further, in the S4, the method of utilizing the gating mechanism to combine the single-hop and two-hop neighbor representations is specifically as follows:

[0028] S41: Utilizing a formula to aggregate the attention weights of the entity and its two-hop neighbors obtained in S2;

[0029] S42: Merging the aggregated representations of the single-hop and two-hop neighbors of the entity through the gating mechanism, controlling the noise introduced by the multi-layer transmission in the S3 graph convolutional neural network, and optimizing the representation of the knowledge graph.

[0030] Further, in the S5, the method of calculating the similarity between entities according to the alignment module to obtain the entity alignment result is specifically as follows:

[0031] S51: Introducing a Manhattan distance calculation formula to calculate the distance measure between the entity embeddings obtained in S4;

[0032] S52: Returning a target entity ranking list of each source entity, and the highest ranking is regarded as alignment.

[0033] A computer system, characterized in that it comprises one or more processors, a computer readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0034] A computer readable storage medium, characterized in that it stores computer executable instructions, which when executed are used to implement the above method.

[0035] Advantages

[0036] The application provides a knowledge graph neighborhood structure non-heterogeneous entity alignment method of fusing multiple social media platforms, collects and fuses data of multiple social media platforms, and constructs a knowledge graph according to user interaction behaviors and theme contents. Since the interaction information and relationships of users on social media are not all of the same architecture, which is reflected as non-heterogeneity of neighborhood structure in the knowledge graph, the graph convolutional neural network is introduced in the application to weaken the influence of the graph structure on the final entity representation. Compared with the traditional graph convolutional network, the GCN obtains the entity single-hop neighbor representation by adding the translation-invariant function, reduces the coupling in the GCN aggregation and translation process, so that the original features of the entity in the translation process are retained instead of only aggregating the entity features of the neighborhood; furthermore, the multi-hop neighbor expansion is introduced in the model to expand the overlap of the entity neighborhood structure, the attention mechanism is used to highlight the useful neighbors of the entity representation and weaken the far neighbors that do not contribute, further optimizing the aggregation effect and reducing the influence of the neighborhood structure non-heterogeneity; then, the noise in the interlayer transmission in the aggregation process is processed, and the gate mechanism is used to control the transmission of the noise. Finally, the similarity between the entity representations is calculated through the alignment module, the specific distance metric between the entity embeddings is used to return a target entity ranking list for each source entity, determine the effect of entity alignment, and obtain the result of entity alignment. The entity alignment model of neighborhood structure non-heterogeneity is used to weaken the feature loss of non-heterogeneous entities in the graph structure conversion and the traditional graph convolutional network conversion, solves the problems caused by the entity non-heterogeneity, and makes the result of entity alignment more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0038] Figure 1 A schematic diagram of the entity alignment structure heterogeneity of the knowledge graph neighborhood structure non-heterogeneity of the application fusing multiple social media platform data;

[0039] Figure 2 A knowledge graph architecture diagram of the entity alignment method of the knowledge graph neighborhood structure non-heterogeneity of the application fusing multiple social media platform data;

[0040] Figure 3 A model framework diagram of the entity alignment method of the knowledge graph neighborhood structure non-heterogeneity of the application fusing multiple social media platform data. DETAILED DESCRIPTION

[0041] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0042] An entity alignment method of a knowledge graph heterogeneous neighborhood structure of a fusion multi-social media platform data, comprising the following steps:

[0043] S1, constructing a knowledge graph according to the interaction and interaction theme content of a user entity on multiple social media platforms;

[0044] S2, calculating the attention weight between entities, and marking neighbor entities that contribute more to the center entity.

[0045] S3, aggregating the single-hop and two-hop neighbors of the entity by a topological invariant function, respectively, and supplementing the single-hop and two-hop neighbor representation information of the entity.

[0046] S4, obtaining the entity representation by using a gating mechanism to combine the single-hop and two-hop neighbor representation.

[0047] S5, obtaining the similarity between entities according to the processed entity representation of the alignment module, and obtaining the final result of alignment.

[0048] The technical solutions of the present application will be described in detail below:

[0049] S1, constructing a knowledge graph according to the interaction and interaction theme content of a user entity on multiple social media platforms;

[0050] In the constructed knowledge graph, the user and the comment theme are used as entity nodes, and the user and the user, the user and the comment theme are used in the form of triplets to form a knowledge graph. For example, users A and B have a social relationship, and A makes a comment on theme 1, which can be converted into triplets <user A, comment, user B> and <user A, comment content, theme 1>.

[0051] S2, calculating the attention weight between entities, and obtaining neighbor entities that contribute more to the center entity.

[0052] The specific steps of the method for obtaining the attention weight between entities are as follows:

[0053] S21, introducing a graph attention network according to the S1 knowledge graph, wherein the network shares two linear transformations in each attention function of the entity:

[0054] The entity representation input into the graph attention network at the lth layer is e ie j ;

[0055] S22, apply a graph attention network to the center entity and the far (two hops and more) neighbor nodes;

[0056] S23, linearly transform the center entity and the neighbor entity using matrix and respectively to obtain the attention weight of the entity;

[0057] The calculation formula of the attention weight of the first layer of is as follows:

[0058]

[0059] S24, normalize the attention weight of the neighbor node of the labeled entity using the softmax function, and the specific softmax function is as follows:

[0060]

[0061] S3, aggregate the single-hop and two-hop neighbors through a topology invariant function to obtain the entity representation.

[0062] The specific steps of aggregating the single-hop and two-hop neighbors through a topology invariant function are as follows:

[0063] S31, according to the knowledge graph and the entity representation processed in S2, extract the feature matrix

[0064] S32, input the feature matrix into the transformation invariant function Mask gate MG (l) , and the calculation formula is as follows:

[0065]

[0066] S33, input the output value MG (l) (H (l) ) of the previous function into the graph convolutional neural network, and the calculation formula is as follows:

[0067]

[0068] S34, interpolate the original features and the linearly transformed features using the transformation invariant function Highway gate, and the calculation formula is as follows:

[0069]

[0070] S35, output the knowledge graph after the entity embedding based on the graph convolutional neural network transformation invariant function, and the calculation formula is as follows:

[0071] H (l+1) = T (l) (H (l) )⊙H (l+1) +(1-T (l) (H (l) ))⊙H (l) ;

[0072] S4, obtain entity representation by using gating mechanism combined with single-hop and two-hop neighbor representation.

[0073] S41, aggregate the attention weight of the entity and its two-hop neighbors obtained by formula S2, and the formula is as follows:

[0074]

[0075] S42, merge the aggregated representation of single-hop and two-hop neighbors of the entity by gating mechanism, control the noise introduced by multi-layer transmission in S3 network, optimize the representation of knowledge graph, and the merging formula is as follows:

[0076]

[0077]

[0078] S5, calculate the similarity between entities according to the alignment module to obtain the result of entity alignment.

[0079] S51, introduce Manhattan distance calculation formula to calculate the distance measure between the entity embeddings obtained by S4;

[0080] The formula for calculating the distance between entity e i1 from knowledge graph KG1 and entity e j2 from knowledge graph KG2 is as follows:

[0081]

[0082] And in this section, the marginal ranking loss function is used, and the formula is as follows:

[0083]

[0084] S52, return the ranking list of target entities of each source entity, and the highest ranking is considered as alignment.

[0085] Finally, the result of entity alignment is obtained.

[0086] The application highlights the entity single-hop neighbor representation obtained by adding the conversion invariant function of the GCN and the screening of the far neighbors by the graph attention mechanism, and improves the entity alignment effect under the non-isomorphic condition of the neighborhood structure. The application provides an end-to-end trained model architecture, and the model does not need to extract features manually, avoids introducing human bias, and reduces the manual design process.

[0087] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application.

Claims

1. A method for entity alignment in the case of non-isomorphic neighborhood structures of knowledge graphs integrating multiple social media platforms, characterized by Here are the steps: S1. Construct a knowledge graph based on the interaction of user entities on multiple social media platforms and the content of the interaction topics; in the knowledge graph constructed by S1, the users and comment topics are both used as entity nodes, and users and users, users and comment topics are represented by triples.<h,r,t> The knowledge graph is constructed in a way as follows; S2. Calculate the attention weights between entities and mark neighboring entities that contribute more to the central entity. In S2, the method for calculating the attention weights between entities is as follows: S21: Based on the knowledge graph, introduce the graph attention network; S22: Apply graph attention network to the central entity and neighboring nodes of two or more hops; S23: Using the Matrix and Perform linear transformation on the central entity and neighbor entities respectively to obtain the attention weight of the entity; S24: Use The function normalizes the attention weights of the labeled entity’s neighbor nodes; S3, aggregate the one-hop and two-hop neighbors of the entity respectively through the topological invariant function to supplement the one-hop and two-hop neighbor representation information of the entity; in S3, the entity is embedded in the knowledge graph through the topological invariant function as follows: S31: Extract the entity feature matrix based on the knowledge graph and the entity representation processed by S2 ; S32: The feature matrix Input conversion invariant function middle; S33: Output the value of the previous function Input graph convolutional neural network; S34: Apply transformation invariant function Interpolate the original features and the linearly transformed features; S35: Output the knowledge graph after entity embedding based on the conversion invariant function of the graph convolutional neural network; S4. Using a gating mechanism to combine the one-hop and two-hop neighbor representations to obtain an entity representation; in S4, the method of using a gating mechanism to combine the one-hop and two-hop neighbor representations is specifically as follows: S41: Aggregate the attention weights of the entity obtained in S2 and its two-hop neighbors using the formula; S42: Merge the aggregated representations of the one-hop and two-hop neighbors of an entity through a gating mechanism to control the noise introduced by multi-layer transmission in the S3 graph convolutional neural network and optimize the knowledge graph representation; S5. Calculate the entity representations after processing using the alignment module to obtain the similarity between entities and obtain the final alignment result. In S5, the method for calculating the similarity between entities using the alignment module to obtain the entity alignment result is specifically as follows: S51: Introduce the Manhattan distance calculation formula to calculate the distance metric between the entity embeddings obtained in S4; S52: Return a sorted list of target entities for each source entity, and the one with the highest ranking is considered aligned.

2. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

3. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images