Publishing house knowledge graph system completion method based on message gating and edge aggregation

Abstract

The invention provides a publishing house knowledge graph system completion method based on message gating and edge aggregation, and belongs to a knowledge graph completion technology. According to the method, neighbor information is independently modeled from node and dual perspectives in a message function stage based on relation sensing gating, and a relation sensing gating mechanism is designed to dynamically adjust the information transmission strength and improve the message discrimination ability; in an information aggregation stage based on multi-head attention, performing weighted aggregation on the double-view-angle information, and generating entity representation by using addition fusion; in the attention-based edge aggregation stage, edge representation is constructed based on interaction of head and tail entities, an edge attention mechanism is designed to perform weighted aggregation on edges belonging to the same relationship, and enhanced relationship representation is generated; and in a decoding stage, an InteractE model is adopted for scoring and inferring, so that high-precision link prediction of the knowledge graph system is realized. Experimental results show that the method is superior to an existing method on multiple reference data sets.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph completion, and relates to a method for completing a publisher's knowledge graph system based on message gating and edge aggregation. Background Technology

[0002] Knowledge graphs are collections of structured knowledge represented by fact triples (head entity, relation, and tail entity). In publishing houses' intelligent systems, they can be used to manage knowledge such as books, authors, journals, and thesaurus, supporting applications such as document recommendation, content retrieval, automatic index generation, and intelligent question answering. However, whether manually or automatically constructed, publishing house knowledge graphs often suffer from incomplete data, which can lead to performance degradation in downstream applications. For example, when some authors' publication information is incomplete or some keywords are missing, the accuracy of document recommendation or content retrieval systems will be affected. Therefore, Knowledge Graph Completion (KGC), as a key task for inferring potential facts and completing missing entities and relations, is of great significance in publishing house knowledge graph management scenarios.

[0003] In recent years, numerous knowledge graph completion methods have employed Knowledge Graph Embedding (KGE) to map entities and relations into representations in a low-dimensional vector space. Traditional embedding models, such as translation-based TransE, tensor decomposition-based DistMult and ComplEx, and convolutional neural network-based ConvE, predict missing triples by modeling known triples. These methods have achieved excellent performance on multiple standard datasets, but they typically ignore the contextual information of entities within the graph structure, making it difficult to fully capture the complex structural dependencies and semantic interactions between entities. This results in limited performance in scenarios involving multi-hop reasoning or high structural noise.

[0004] With the widespread application of Graph Neural Networks (GNNs) in graph-structured data modeling, more and more research is introducing them into knowledge graph completion tasks to more effectively integrate complex relational structures and multi-level semantic information between entities. GNN methods typically employ an encoder-decoder framework, where the encoder uses the GNN to model the contextual representation of each entity, and the decoder predicts scores for candidate triples based on the entity and relation representations. In this process, the GNN passes information from an entity's neighboring nodes and their relationships through message passing mechanisms and generates context-aware entity representations through iterative aggregation and update operations. Compared to traditional embedding models, GNNs can model structural dependencies and relational semantics more deeply, thereby improving the reasoning ability for missing knowledge in complex graph structures. Numerous studies have shown that graph neural networks significantly improve the accuracy and robustness of knowledge graph completion in triple prediction tasks. Typical examples include CompGCN, which integrates relation and entity features; RAGAT, which applies attention mechanisms to heterogeneous graphs; and SE-GNN, which introduces triple semantic information. These methods provide effective paths for more efficient use of graph structures and drive the continuous improvement of the performance of knowledge graph completion tasks.

[0005] Although graph neural network methods have made significant progress in knowledge graph completion tasks in recent years, existing models still have two key problems in knowledge graph system completion tasks for publishing houses: (1) lack of relation mapping patterns and semantic role-aware message passing mechanisms. Although current mainstream methods introduce attention mechanisms in the aggregation stage to assign weights to different neighbor nodes and their connections, thereby achieving a certain degree of differentiated modeling, attention is essentially a relative ranking of the importance of neighbor information. It only adjusts the contribution ratio of different messages during aggregation and does not model the message function based on relation type. However, in knowledge graphs, relations can usually be divided into four types according to the mapping characteristics between head entities and tail entities: 1-1, 1-N, N-1, and NN. These relation mapping patterns not only have significant differences in structural distribution, but also have different effects on the representation of target entities during message passing. Specifically, in triples During message passing, when relation r belongs to type 1-1, 1-N, or NN, the header entity... The semantic information carried often helps to enhance tail entities. The quality of the representation. Especially in 1-N types, the head entity This can typically provide more discriminative semantic information, thereby improving the model's ability to identify tail entities. The ability to distinguish. Conversely, when the relationship When it belongs to type N-1, multiple head entities point to the same tail entity. The information passed in by neighboring nodes may be redundant or even noisy, which may interfere with the representation learning of tail entities. Therefore, the lack of a message passing mechanism that is aware of relation types may lead to excessive averaging of information transmission, thereby weakening the semantic expression ability of the model. (2) Insufficient relation embedding modeling and aggregation mechanism. Although the aggregation and representation learning of entity features have received sufficient attention in current knowledge graph embedding models, most methods still adopt a relatively simple strategy for relation embedding modeling. The typical approach is to perform linear transformation or fixed mapping on relations, without making full use of the contextual semantics and structural information carried by relations as edges. In the publishing house knowledge graph system, relations are not only channels connecting entities, but also carry clear semantic constraints, type guidance, and contextual features formed by the interaction of multiple pairs of head and tail entities. If there is a lack of dynamic modeling and aggregation mechanism for relation semantics, it is difficult to capture the semantic changes of the same relation in different contexts, and it is also impossible to effectively utilize this semantic difference in the reasoning process, thereby limiting the performance of the model in the publishing house knowledge graph system completion task. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method for completing a publisher's knowledge graph system based on message gating and edge aggregation. In the message function stage, based on relation-aware gating, message functions are modeled independently from both node and dual perspectives, and a relation-aware gating mechanism is designed to effectively adjust the information transmission intensity under different relation mapping patterns and semantic roles. In the information aggregation stage based on multi-head attention, information from both node and dual perspectives is independently weighted and aggregated, and an entity embedding representation containing dual-perspective semantic features is obtained through additive fusion. In the edge aggregation stage based on attention, edge embedding representations are constructed based on the semantic interaction of head and tail entities in triples, and an edge attention mechanism is designed to weight and aggregate all edges belonging to the same relation, thereby obtaining an enhanced relation embedding representation. In the decoding stage, the InteractE model is used for scoring and inference, achieving high-precision link prediction for the publisher's knowledge graph system.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A knowledge graph completion method for publishers based on message gating and edge aggregation is proposed. It independently models message functions from both node-level and dual-level perspectives, and designs a relation-aware gating mechanism to enhance the discriminative power of neighbor messages. Based on the semantic interaction between head and tail entities, it constructs an edge embedding representation and designs an edge attention mechanism to weighted aggregate all edges belonging to the same relation, thereby obtaining an enhanced relation embedding representation. The method includes the following steps:

[0009] Step 1: Design message functions based on relation-aware gating

[0010] This step independently models the structural information of neighboring nodes and the semantic interaction information between nodes and relations in the publisher's knowledge graph system from both node-level and dual-level perspectives. It also designs a relation-aware gating mechanism to dynamically adjust the transmission strength of message functions under different relation mapping modes and semantic roles, thereby improving the discriminative power of message representation.

[0011] Step 2: Design an information aggregation mechanism based on multi-head attention.

[0012] After obtaining the dual-view message function filtered by the relation-aware gating mechanism, this step introduces a multi-head attention mechanism to independently weight and aggregate neighbor information in the publisher's knowledge graph system from both node-level and dual-level perspectives. Finally, the final entity representation containing dual-view semantic features is obtained through additive fusion.

[0013] Step 3: Design an attention-based edge aggregation mechanism

[0014] This step constructs an edge embedding representation based on the semantic interaction of the head and tail entities of triples, and designs an edge attention mechanism to weighted aggregate all edges belonging to the same relation, generating a relation embedding representation with stronger semantic consistency. To improve the model's ability to model low-frequency relations, residual connections are used to fuse the original relation embedding with the enhanced representation, thereby enhancing the robustness and generalization ability of the relation embedding representation.

[0015] Step 4: Design an InteractE-based decoder

[0016] This step uses the InteractE model as a decoder to score candidate tail entities based on the embedding representations of head entities and relations, aiming to accurately identify and fill query triples in the publisher's knowledge graph system. Missing parameters (Tail entity). During model training, the binary cross-entropy loss function is used for optimization, which effectively distinguishes between positive and negative samples, improves prediction accuracy and generalization ability, and enables efficient inference of missing connections in the publisher's knowledge graph system.

[0017] The beneficial effects of this invention are demonstrated by experiments on two public datasets, FB15k-237 and WN18RR, showing that this method achieves significant improvements over existing state-of-the-art models across multiple evaluation metrics. This is attributed to improvements in message passing difference modeling and attention-based edge aggregation, which effectively enhance the overall expressive power and inference performance of the model. Experimental results show that this invention outperforms existing methods on multiple benchmark datasets, effectively improving the accuracy of knowledge graph system completion for publishers and providing reliable technical support for downstream applications such as document recommendation, intelligent retrieval, and content indexing. Attached Figure Description

[0018] Figure 1 This invention provides a method for completing a publisher's knowledge graph system based on message gating and edge aggregation. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] The overall process diagram of the present invention is as follows: Figure 1 As shown, a method for completing a publisher's knowledge graph system based on message gating and edge aggregation includes the following steps:

[0021] Step 1: Design message functions based on relation-aware gating

[0022] This invention uses message functions to represent how the central node of a publisher's knowledge graph system obtains information from neighboring nodes and their connections. Message functions are independently modeled for neighboring information from both node-level and dual-level perspectives, and a relationship-aware gating mechanism is designed to dynamically adjust the intensity of information transmission.

[0023] Step 1.1: Node-level message function

[0024] Node-level message functions are used to model the semantic and structural features of neighboring nodes in a publisher's knowledge graph system. If the entity... Based on relationships Connect to entity Then their node-level message functions The definition is as follows:

[0025]

[0026] in, The initial embedding of the entity and the weight of the relation type. The definition is as follows:

[0027]

[0028] in, This is the initial dimension for entity embedding; Embed dimensions for hidden layers; It is the set of all positive relations in the publisher's knowledge graph. It is by The set consisting of the inverse relations of each relation in the set. Used to represent self-loop relationships, i.e., edges from an entity to itself. A unary set containing only self-loop relations. This is the transformation matrix corresponding to the positive relationship. This is the transformation matrix corresponding to the inverse relationship. This is the transformation matrix corresponding to the self-loop relation.

[0029] Step 1.2: Dual-level message function

[0030] Dual-level message functions are used to model the semantic interaction features of neighboring nodes and their connections in a publisher's knowledge graph system. If entities Based on relationships Connect to entity Then their dual-level message functions The calculation is as follows:

[0031]

[0032] in, The definition is shown in step 1.1. Indicates the initial embedding of the entity. Represents the initial embedding of the relation. It is a composite operator, and the specific calculation method is as follows:

[0033]

[0034] in, It is the learnable matrix corresponding to the relation. This represents element-wise multiplication (Hadamard product).

[0035] Step 1.3: Relationship-Aware Message Gating

[0036] In the process of completing a publisher's knowledge graph system, the contributions of different mapping patterns and semantic roles to information dissemination are uneven. If a uniform message passing strategy is adopted, the model often struggles to effectively distinguish the amount of information carried by different relationships, resulting in limited entity representation capabilities. For example, when maintaining a 1-N relationship of "author-writing-book" in the system, such as triples (author A, writing, book B) and (author A, writing, book C), where the same author A writes multiple books, the information from the head entity "author" is particularly important for the representation of "book" in the system because the head entity can provide more discriminative messages. Therefore, a larger gating coefficient should be applied to enhance information flow. Conversely, in N-1 relationships such as "author-belonging to-publisher," such as triples (author A, belonging to, publisher C) and (author B, belonging to, publisher C), multiple author entities point to the same publisher entity. In this case, if the incoming messages are not controlled, redundant or irrelevant semantics may be introduced, affecting the accuracy of the representation of "publisher C." Therefore, the model needs to be assigned a smaller gating coefficient to reduce noise interference. To address this, this invention proposes a relation-aware gating mechanism to dynamically adjust the transmission strength of neighbor messages under different relations in the system. Specifically, for any triplet edge... ,relation The embedding is The following gating strategy is used to filter and control the triple message:

[0037]

[0038] in, For relationship The gating coefficient is used to measure the importance of this relationship in the system completion process. It is a linear transformation matrix. For bias terms, The Sigmoid activation function is used to map the gate value to... The interval is used to represent the importance weight of information under different relationships. Finally, a gating coefficient is applied to the original message function to achieve weighted control of neighbor information:

[0039]

[0040]

[0041] in, and These represent the node-level message function and the dual-level message function, respectively, after being weighted by relation-aware gating coefficients. Through the relation-aware gating mechanism, the publisher's knowledge graph system can adaptively adjust the information transmission strength according to the semantic contribution of relation types, thereby effectively mitigating the interference caused by redundant relations and improving the discriminative power of neighbor messages.

[0042] Step 2: Design an information aggregation mechanism based on multi-head attention

[0043] After obtaining the dual-perspective messages passed by the message function based on relation-aware gating, this invention further introduces a multi-head attention mechanism to perform weighted aggregation of neighbor information in the publisher's knowledge graph system from both node-level and dual-level perspectives, and obtains the final entity representation containing multi-perspective semantics through additive fusion.

[0044] Step 2.1: Node-level attention information aggregation

[0045] Node-level attention aims to determine which node-level messages contribute more significantly to the update of the central node representation in a publisher's knowledge graph system. This mechanism primarily focuses on the semantic and structural importance of each neighbor's information to the central node representation. The attention weights for node-level messages are defined as follows:

[0046]

[0047] in, Indicates from entity Through relationships Passed to node Node-level messages The absolute attention coefficient. It is a learnable attention coefficient. To obtain a relative attention value, for... Perform softmax normalization:

[0048]

[0049] in, It is the attention coefficient of node-level messages. The target node Neighboring entities, It is a connecting entity and entity A set of relationships, It is an entity Through relationships Passed to node The absolute attention coefficient of node-level messages. Ultimately, the node... Node-level embedding It can be represented as:

[0050]

[0051] in, It is an activation function, specifically used... ; Represents nodes The set of all connected neighboring entities and their relationship pairs.

[0052] Multi-head attention has been widely used in knowledge graph embedding to stabilize the training process and improve model performance. This invention introduces a multi-head attention mechanism to enhance the expressive power of entities in a publisher's knowledge graph system. Formalized as follows:

[0053]

[0054] in, The number of attention heads.

[0055] Step 2.2: Dual-level attention information aggregation

[0056] Dual-level attention aims to determine which dual-level information contributes more to the update of the central node representation in a publisher's knowledge graph system. This mechanism captures the importance of different relation types, allowing the model to assign different weights to different relations. This ensures that the model focuses on the most relevant relations, which might be ignored in node-level attention mechanisms. The attention mechanism is introduced to calculate its dual-level attention weights to the central node:

[0057]

[0058] in, Indicates from entity Through relationships Passed to node Dual messages The absolute attention coefficient. It is a learnable attention coefficient. To obtain a relative attention value, for... Perform softmax normalization:

[0059]

[0060] in, For attention parameters in the dual-level path, The target node Neighboring entities, It is a connecting entity and entity A set of relationships. Ultimately, nodes. Dual-level embedding It can be represented as:

[0061]

[0062] in, It is an activation function, specifically used... Similarly, this step uses a multi-head attention mechanism to enhance the expressive power of entities in the publisher's knowledge graph system:

[0063]

[0064] Step 2.3: Dual-view information fusion

[0065] The two paths described above model neighbor information from different perspectives and capture graph structure information at different levels. The embeddings obtained from the two perspectives are then added together and fused to finally obtain the node. Embedded express:

[0066]

[0067] in, For nodes The final output represents, For nodes Node-level embedding representation, For nodes The dual-level embedding representation.

[0068] Step 3: Design an attention-based edge aggregation mechanism

[0069] To enhance the expressive power of relation embedding in a publisher's knowledge graph system, this invention proposes an attention-based edge aggregation mechanism. Specifically, for each relation... (e.g., "author-write-book", "publisher-publish-book", "book-belongs to-category"), count all edge sets connecting this relationship based on the triple, denoted as... , among which the side Representing entities With entity Through relationships Connected. First, based on the core assumptions of the classic knowledge graph embedding model TransE. Constructing relational representation as That is, the relation vector can be viewed as a translation vector from the head entity embedding to the tail entity embedding, thereby effectively capturing the semantic information and structural features of the relation. For each triple... Embedding of its edges It is expressed as follows:

[0070]

[0071] in, and These are learnable linear transformation matrices used to transform the tail and head entities, respectively, thereby improving the edge embedding representation capability of the publisher's knowledge graph system. To obtain the relations... The overall embedding representation requires aggregating all edge embeddings belonging to the relation. However, simply averaging or summing these embeddings often fails to adequately reflect the semantic differences in importance between different edges, resulting in insufficient relation representation. To address this issue, this step designs an edge attention mechanism, dynamically adjusting the contribution weight of each edge embedding to the relation representation to improve the aggregation effect. Specifically, a single-layer feedforward neural network with shared parameters is used to score the importance of each edge embedding. :

[0072]

[0073] in, It is a linear transformation matrix. It's an activation function used to enhance nonlinear modeling capabilities. Next, all edges are scored. After performing softmax normalization, we obtain its relative attention coefficient:

[0074]

[0075] in, Represents the head entity of the edge. Represents the tail entity of the edge. Indicates from entity To the entity Score the attention level on the side. Indicates a membership relationship The set of all related edges. Finally, the relation. Context-enhanced embedding representation By embedding all edges According to attention weight Weighted aggregation yields:

[0076]

[0077] in, For relationship The number of associated edges, This represents a non-linear activation function, specifically used... Since different relationships have different numbers of edges, an averaging method is used to integrate the edge representations to prevent the deviation in the number of edges under different relationship types from affecting the model.

[0078] However, in publisher knowledge graph systems, edges for certain relation types are sparse, and simply relying on attention aggregation of context entities may be insufficient to capture their semantic information, leading to inadequate relation representation learning. To address this, a residual connection mechanism is introduced to embed the initial relation... This is integrated into context-based augmented representations to improve the ability to model sparse relations. The specific computational form is as follows:

[0079]

[0080] in, For relationship The final output indicates that For relationship The context-enhanced embedding representation obtained by aggregation through edge attention mechanism. For relationship The initial embedding representation, It is a linear transformation matrix. This residual structure can enhance contextual modeling capabilities while preserving the original semantic information.

[0081] Step 4: Design an InteractE-based decoder

[0082] After entity and relation encoding, the model obtains the entity embedding output representation from the publisher's knowledge graph system. Relational embedding output representation To complete the knowledge graph system for publishing houses, this invention uses InteractE as a decoder to score potential triples (e.g., predicting "books a certain author might write" or "books a certain publishing house might publish"). InteractE fully explores the interaction patterns between entities and relationships through three mechanisms: feature permutation, checkerboard feature rearrangement, and recurrent convolution. For input pairs... InteractE first generates Random permutations:

[0083]

[0084] in, To represent the entity-relation pairs generated from the input entities A sequence of random permutations is used to enhance feature diversity and capture different interaction patterns. Then, for each permutation pair... Apply chessboard-style rearrangement operation ,Right now:

[0085]

[0086] InteractE for triples The probability estimate can be formally represented as follows:

[0087]

[0088] in, The model predicts triples. The probability that this is true. This represents flattening a tensor into a vector, and ⊛ represents depthwise convolution. For convolution kernel parameters, It is the weight matrix. (Function) and These represent the ReLU activation function and the Sigmoid activation function, respectively. During training, this invention employs the binary cross-entropy loss function. The (BinaryCross-Entropy Loss) is optimized and defined as follows:

[0089]

[0090] in, The total number of candidate tail entities. It is a triplet The corresponding labels are 1 to indicate that the triplet actually exists, and 0 to indicate that the triplet does not exist. These are the corresponding scores. This invention employs a 1-Vs-All training strategy, that is, for each head entity-relation pair... Simultaneously, all candidate tail entities are scored, and a loss function is constructed based on global samples, thereby effectively improving training efficiency and model generalization ability. Through the above design, the model of this invention can simultaneously take into account the structural information modeling of entities and relationships and target prediction, effectively improving the overall performance of the knowledge graph system completion task for publishers.

[0091] This invention presents extensive knowledge graph completion experiments on two real-world datasets, FB15k-237 and WN18RR. The performance of the model is evaluated using the mean reciprocal ranking (MRR) and the top N hit rate (Hits@N, N=1,3,10). Results show that this method significantly outperforms all state-of-the-art methods on both benchmark datasets and achieves optimal results on most evaluation metrics. For example, using the mean reciprocal ranking, which best reflects model performance, this method shows improvements of 1.64% and 1.63% compared to the second-best result on FB15k-237 and WN18RR, respectively, indicating that this method can more accurately predict truly missing entities.

[0092] Table 1 shows the knowledge graph completion results based on the two datasets.

[0093]

Claims

1. A method for completing a publisher's knowledge graph system based on message gating and edge aggregation, characterized in that, Includes the following steps: Step 1: Design message functions based on relation-aware gating Message functions are used to represent how the central node of the publisher's knowledge graph system obtains information from neighboring nodes and their connections. Message functions are independently modeled for neighboring information from both node-level and dual-level perspectives. A relationship-aware gating mechanism is designed to dynamically adjust the intensity of information transmission. Step 2: Design an information aggregation mechanism based on multi-head attention. After obtaining the dual-perspective neighbor information passed in step 1, a multi-head attention mechanism is introduced to perform weighted aggregation of neighbor information in the publisher's knowledge graph system from both node-level and dual-level perspectives, and obtain the final entity representation containing multi-perspective semantics through additive fusion. Step 3: Design an attention-based edge aggregation mechanism An edge embedding representation is constructed based on the semantic interaction of the head and tail entities of triples. An edge attention mechanism is designed to perform weighted aggregation on all edges belonging to the same relation, generating a relation embedding representation with stronger semantic consistency. A residual connection is used to fuse the initial relation embedding with the enhanced representation, thereby enhancing the robustness and generalization ability of the relation embedding representation. Step 4: Design an InteractE-based decoder Using the InteractE model as the decoder, candidate tail entities are scored based on the embedding representations of head entities and relations, aiming to accurately identify and fill query triples in the publisher's knowledge graph system. Missing parameters .

2. The method for completing a publisher's knowledge graph system based on message gating and edge aggregation as described in claim 1, characterized in that, Step 1 is as follows: Step 1.1: Node-level message function Node-level message functions are used to model the semantic and structural features of neighboring nodes in a publisher's knowledge graph system; if the entity Based on relationships Connect to entity Then their node-level message functions The definition is as follows: ； in, Indicates entity embedding and relation type weight. The definition is as follows: ； in, This is the initial dimension for entity embedding; Embed dimensions for hidden layers; It is the set of all positive relations in the publisher's knowledge graph. It is by The set consisting of the inverse relations of each relation in the set. Used to represent self-loop relationships, i.e., edges from an entity to itself. A unary set containing only self-loop relations. This is the transformation matrix corresponding to the positive relationship. This is the transformation matrix corresponding to the inverse relationship. This is the transformation matrix corresponding to the self-loop relation; Step 1.2: Dual-level message function Dual-level message functions are used to model the semantic interaction features of neighboring nodes and their connections in a publisher's knowledge graph system; if entities Based on relationships Connect to entity Then their dual-level message functions The calculation is as follows: ； in, The definition is shown in step 1.

1. Indicates the initial embedding of the entity. Represents the initial embedding of the relation. It is a composite operator, and the specific calculation method is as follows: ； in, It is the learnable matrix corresponding to the relation. This represents element-wise multiplication (Hadamard product). Step 1.3: Relationship-Aware Gating Mechanism For any triple edge ,relation The embedding is The following gating strategy is used to filter and control the triple message: ； in, For relationship The gating coefficient, It is a linear transformation matrix. For bias terms, The Sigmoid activation function is used to map the gate value to... The interval is then used; finally, a gating coefficient is applied to the original message function to achieve weighted control of neighbor information. ； ； in, and These represent the node-level message function and the dual-level message function, respectively, after being weighted by message gating coefficients.

3. The method for completing a publisher's knowledge graph system based on message gating and edge aggregation according to claim 1, characterized in that, Step 2 is as follows: Step 2.1: Node-level attention information aggregation Define the attention weights for node-level messages as follows: ； in, Indicates from entity Through relationships Passed to node Node-level messages The absolute attention coefficient; It is a learnable attention coefficient; in order to obtain a relative attention value, for Perform softmax normalization: ； in, It is the attention coefficient of node-level messages. The target node Neighboring entities, It is a connecting entity and entity A set of relationships, It is an entity Through relationships Passed to node The absolute attention coefficient of node-level messages; ultimately, the node Node-level embedding Represented as: ； in, It is an activation function; Represents nodes The set of all connected neighboring entities and their relationship pairs; A multi-head attention mechanism is introduced to enhance the expressive power of entities in the publisher's knowledge graph system; formally represented as follows: ； in, For the number of heads; Step 2.2: Dual-level attention information aggregation Introduce an attention mechanism to compute the dual-level attention weights for the central node: ； in, Indicates from entity Through relationships Passed to node Border-level messages The absolute attention coefficient; It is a learnable attention coefficient; in order to obtain a relative attention value, for Perform softmax normalization: ； in, For attention parameters in the dual-level path, The target node Neighboring entities, It is a connecting entity and entity A set of relationships; ultimately, nodes Edge embedding Represented as: ； in, It is an activation function; Accordingly, a multi-head attention mechanism is used to enhance the expressive power of entities in the publisher's knowledge graph system: ； Step 2.3: Dual-view information fusion The embeddings obtained from the two perspectives are additively fused to obtain the final node. Embedded representation : ； in, For nodes The final output represents, For nodes Node-level embedding representation, For nodes The dual-level embedding representation.

4. The method for completing a publisher's knowledge graph system based on message gating and edge aggregation as described in claim 1, characterized in that, Step 3 is as follows: For each type of relationship The set of all edges connecting this relation is counted based on the triples, denoted as . , among which the side Representing entities With entity Through relationships Connected; First, based on the assumptions of the knowledge graph embedding model TransE Constructing relational representation as For each triple Embedding of its edges It is expressed as follows: ； in, and These are learnable linear transformation matrices, used to transform the tail entity and the head entity, respectively. An edge attention mechanism is designed to dynamically adjust the contribution weight of each edge embedding to the relation representation; specifically, a single-layer feedforward neural network with shared parameters is used to score the importance of each edge embedding. : ； in, It is a linear transformation matrix. It's an activation function; next, it scores all the edges. After performing softmax normalization, we obtain its relative attention coefficient: ； in, Represents the head entity of the edge. Represents the tail entity of the edge. Indicates from entity To the entity Score the attention level on the side. Indicates a membership relationship The set of all related edges; finally, the relation Context-enhanced embedding representation By embedding all edges According to attention weight Weighted aggregation yields: ； in, For relationship The number of associated edges, Represents a nonlinear activation function; Introducing a residual join mechanism to initially embed the relation When integrated into the context-enhanced embedding representation, the specific computational form is as follows: ； in, For relationship The final output indicates that For relationship The context-enhanced embedding representation obtained by aggregation through edge attention mechanism. For relationship The initial embedding representation, It is a linear transformation matrix.

5. The method for completing a publisher's knowledge graph system based on message gating and edge aggregation according to claim 1, characterized in that, Step 4 is as follows: InteractE is used as the decoder to score latent triples; for input pairs InteractE first generates Random permutations: ； in, To represent the entity-relation pairs generated from the input entities A sequence of random permutations; then, for each permutation pair... Apply chessboard-style rearrangement operation ,Right now: ； InteractE for triples The probability estimate is formally represented as follows: ； in, The model predicts triples. The probability that this is true. This represents flattening a tensor into a vector, and ⊛ represents depthwise convolution. For convolution kernel parameters, It is a weight matrix; function and These represent the ReLU activation function and the Sigmoid activation function, respectively.

6. The method for completing a publisher's knowledge graph system based on message gating and edge aggregation according to claim 5, characterized in that, During training, a binary cross-entropy loss function is used for optimization, defined as follows: ； in, The total number of candidate tail entities. It is a triplet The corresponding label 1 indicates that the triple actually exists, i.e., it is the head entity. ,relation Tail entity A valid triplet is formed; 0 indicates that the triplet does not exist. These are the corresponding scores; a 1-Vs-All training strategy is used, meaning for each head entity-relation pair... At the same time, all candidate tail entities are scored, and a loss function is constructed based on global samples.

Images