Recommendation method and device based on knowledge graph, equipment and storage medium

Abstract

The invention relates to a recommendation method based on a knowledge graph, and the method comprises the steps: carrying out the semantic analysis of knowledge data of a target domain through a bidirectional LSTM auto-encoder, and generating a semantic vector; in combination with a semantic vector and named entity recognition technology and a semantic role labeling technology, entities and relationships thereof are recognized, a triple is constructed, and a structured knowledge graph is generated. Node feature extraction and embedded representation are performed in a knowledge graph, an initial graph neural network model is constructed in combination with graph structure information, a target graph neural network model is generated through training and optimization, nodes are scored and sorted, and a target object is recommended. According to the invention, the accuracy and intelligent level of recommendation can be improved, the personalization and matching degree of recommendation results are enhanced, and the problem of inaccurate recommendation results in the prior art is significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a recommendation method, device, equipment and storage medium based on knowledge graph. Background Art

[0002] In the field of talent recommendation, existing systems mostly rely on resume keyword matching or rule-based algorithms. This method is difficult to fully explore the potential and suitability of candidates, performs poorly in personalized recommendations, and cannot fully consider the multi-dimensional characteristics of candidates (such as skills, experience, cultural fit, etc.). Due to the lack of deep semantic analysis and relationship recognition capabilities, the system can often only provide recommendation results based on surface features, which is difficult to meet the precise needs of financial institutions for high-end talents.

[0003] Although natural language processing (NLP) technology has been applied to the field of talent recommendation to a certain extent, the existing semantic analysis models still have the problem of insufficient accuracy. When processing long texts or complex contexts, these models often cannot accurately identify subtle differences in candidate information, resulting in biased information interpretation and affecting the effectiveness of talent recommendation.

[0004] Current technologies still face challenges in integrating and identifying relationships between multi-source heterogeneous data when constructing knowledge graphs. Traditional methods are difficult to effectively extract and integrate complex relationships between candidates (such as project participation, cooperation experience, etc.), resulting in deficiencies in the accuracy and comprehensiveness of the constructed knowledge graphs.

[0005] Although graph neural networks (GNNs) have shown great potential in processing graph-structured data, their application in talent recommendation systems still faces many limitations. The computational complexity of GNN models is relatively high, especially when processing large-scale knowledge graphs, which requires a long model training time and consumes a lot of computing resources. In addition, the model design of GNNs often needs to be customized according to specific application scenarios, which increases the difficulty of development and deployment.

[0006] These problems limit the overall performance of the talent recommendation system and make it difficult to provide efficient and accurate recommendation results for financial institutions. Therefore, it is urgent to solve these key problems through technological innovation and improve the intelligence and personalization level of the talent recommendation system. Summary of the invention

[0007] The main purpose of the present invention is to provide a recommendation method, device, equipment and storage medium based on knowledge graph, aiming to solve the technical problems in existing recommendation technologies, such as the lack of effective semantic analysis, relationship recognition and complex network structure processing capabilities, resulting in insufficient accuracy and personalization of recommendation results.

[0008] To achieve the above object, the present invention provides a recommendation method based on knowledge graph, comprising:

[0009] Acquire knowledge data in a target domain, perform semantic analysis on the knowledge data through a bidirectional LSTM autoencoder, and generate a semantic vector of the knowledge data;

[0010] Combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, identifying entities and relationships between entities from the knowledge data, and constructing a triple containing a subject, an object and a predicate;

[0011] Building a structured knowledge graph based on the triples, extracting node features and embedding representation in the knowledge graph, and generating a feature vector for each node;

[0012] Combine the feature vector of each node with the graph structure information in the knowledge graph to build the initial graph neural network model;

[0013] Training and optimizing the initial graph neural network model to generate a target graph neural network model;

[0014] Based on the target graph neural network model, the nodes representing entities in the target domain are scored and sorted, and the nodes that meet the recommendation conditions are recommended as target objects in the target domain.

[0015] In one embodiment, the knowledge data is semantically analyzed by a bidirectional LSTM autoencoder to generate a semantic vector of the knowledge data, including:

[0016] Perform text segmentation and word embedding processing on knowledge data to convert text into word vectors of fixed dimensions;

[0017] Inputting the word vector into the bidirectional LSTM autoencoder, wherein the bidirectional LSTM autoencoder includes a forward LSTM network and a reverse LSTM network;

[0018] The word vectors are processed from the beginning to the end of the sentence through the forward LSTM network, and the word vectors are processed from the end to the beginning of the sentence through the reverse LSTM network to generate the forward hidden state and the reverse hidden state respectively;

[0019] At each time step, the forward hidden state and the reverse hidden state are concatenated to form a bidirectional hidden state that contains contextual information.

[0020] Apply activation function to the concatenated bidirectional hidden state, and generate a global semantic vector through concatenation or pooling of the processed output;

[0021] The generated global semantic vector is input into the decoder, and the original input sentence is reconstructed through the decoder, and the difference between the decoder output and the original sentence is calculated as the loss value;

[0022] The model parameters are adjusted according to the loss value through the back-propagation algorithm to optimize the semantic representation ability of the model, and finally a semantic vector reflecting the semantics of the knowledge data is generated.

[0023] In one embodiment, the generated semantic vector, named entity recognition technology and semantic role labeling technology are combined to identify entities and relationships between entities from the knowledge data, and construct triples containing subject, object and predicate, including:

[0024] Based on the generated semantic vectors, a hierarchical clustering algorithm is used to cluster the semantic vectors representing entity names to obtain multiple name clusters;

[0025] The named entity recognition technology is used to perform entity recognition on the clustered name clusters to identify entities representing the same actual objects;

[0026] Based on the generated semantic vectors, a hierarchical clustering algorithm is used to cluster the semantic vectors representing the relationship between entities to obtain multiple relationship clusters;

[0027] The semantic role labeling technology is used to identify the relationship clusters after clustering, and the relationship sentences expressing the same semantic relationship or having the same grammatical structure are classified into a unified relationship type;

[0028] Based on the clustered entities and the unified relationship type, a triple consisting of subject, object and predicate is constructed.

[0029] In one embodiment, node feature extraction and embedding representation are performed in the knowledge graph to generate a feature vector for each node, including:

[0030] Extracting basic attributes of each node in the knowledge graph, wherein the basic attributes include node type, number of connected edges, node degree, and information of neighboring nodes;

[0031] Set the step size and transfer probability of the random walk, and use the random walk technology to perform multiple random walk operations on each node to generate multiple node sequences;

[0032] Using the word2vec method, the generated node sequence is trained through the skip-gram model or the CBOW model to generate the feature vector of each node.

[0033] In one embodiment, the feature vector of each node is combined with the graph structure information in the knowledge graph to construct an initial graph neural network model, including:

[0034] Multiply the feature vector of each node by the pre-trained weight matrix to generate a preliminary score value for the node, and adjust the preliminary score value based on the graph structure information in the knowledge graph;

[0035] Calculate the weight value of each node, and perform nonlinear processing through the activation function to obtain the weight value after activation;

[0036] Normalize the activated weight values ​​so that the sum of the weight values ​​of each node is 1, and generate a normalized weight vector;

[0037] Using the normalized weight vector and the node's in-degree, the node's preliminary score value is weighted and adjusted to generate the node's final feature representation;

[0038] In each layer of the graph neural network, the final feature representation of the node and the normalized weight vector are used to control the intensity of information propagation, and information is aggregated and updated through the network layer to finally build a complete initial graph neural network model.

[0039] In one embodiment, the initial graph neural network model is trained and optimized to generate a target graph neural network model, including:

[0040] Use mean square error as the loss function to calculate the prediction error of the initial graph neural network model on the training set;

[0041] Using the calculated prediction error, the model parameters are adjusted using the Adam optimizer through the back-propagation algorithm to minimize the loss function.

[0042] Perform multiple iterations of training, and use the validation set to evaluate and adjust the model performance after each training round.

[0043] After multiple rounds of training and optimization, the optimized target graph neural network model is finally generated for prediction and recommendation tasks.

[0044] In one embodiment, based on the target graph neural network model, scoring and ranking nodes representing entities in the target domain include:

[0045] Extracting a final feature representation of each node from the target graph neural network model, wherein the final feature representation is a vector generated through a multi-layer network propagation and information aggregation process;

[0046] Calculate the centrality index of the relative importance of each node in the graph structure through degree centrality, closeness centrality, betweenness centrality or eigenvector centrality algorithms;

[0047] Using the preset transformation function, the feature vector of each node is combined with the centrality index to generate a preliminary evaluation value for each node;

[0048] Analyze the connection strength between the node and its neighboring nodes, and adjust the preliminary evaluation value of each node based on the analysis results to generate the final evaluation value;

[0049] The nodes representing entities in the target domain are ranked based on the final evaluation value of each node.

[0050] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and a knowledge graph-based recommendation program stored in the memory and executable on the processor, wherein the knowledge graph-based recommendation program, when executed by the processor, implements the steps of the knowledge graph-based recommendation method as described above.

[0051] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer storage medium, on which a recommendation program based on a knowledge graph is stored, and when the recommendation program based on the knowledge graph is executed by a processor, the steps of the recommendation method based on the knowledge graph as described above are implemented.

[0052] Beneficial effects: The present invention relates to a recommendation method based on knowledge graph, which uses a bidirectional LSTM autoencoder to perform semantic analysis on the knowledge data in the target field to generate a semantic vector. The semantic vector is combined with the named entity recognition technology and the semantic role labeling technology to identify entities and their relationships, construct triples and generate a structured knowledge graph. Node feature extraction and embedding representation are performed in the knowledge graph, and an initial graph neural network model is constructed in combination with the graph structure information. The target graph neural network model is generated through training and optimization, and the nodes are scored and sorted to recommend the target object. The present invention can improve the accuracy and intelligence level of recommendations, enhance the personalization and matching of recommendation results, and significantly improve the problem of inaccurate recommendation results in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0054] Figure 1 This is a flow chart of an embodiment of a recommendation method based on knowledge graph of the present invention;

[0055] Figure 2 This is a functional module diagram of a preferred embodiment of a recommendation device based on knowledge graph of the present invention;

[0056] Figure 3 It is a structural diagram of the device hardware operating environment involved in the computer device embodiment of the present invention. DETAILED DESCRIPTION

[0057] 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.

[0058] In the field of talent recommendation, existing systems mostly rely on resume keyword matching or rule-based algorithms. This method is difficult to fully explore the potential and suitability of candidates, performs poorly in personalized recommendations, and cannot fully consider the multi-dimensional characteristics of candidates (such as skills, experience, cultural fit, etc.). Due to the lack of deep semantic analysis and relationship recognition capabilities, the system can often only provide recommendation results based on surface features, which is difficult to meet the precise needs of financial institutions for high-end talents.

[0059] Although natural language processing (NLP) technology has been applied to the field of talent recommendation to a certain extent, the existing semantic analysis models still have the problem of insufficient accuracy. When processing long texts or complex contexts, these models often cannot accurately identify subtle differences in candidate information, resulting in biased information interpretation and affecting the effectiveness of talent recommendation.

[0060] Current technologies still face challenges in integrating and identifying relationships between multi-source heterogeneous data when constructing knowledge graphs. Traditional methods are difficult to effectively extract and integrate complex relationships between candidates (such as project participation, cooperation experience, etc.), resulting in deficiencies in the accuracy and comprehensiveness of the constructed knowledge graphs.

[0061] Although graph neural networks (GNNs) have shown great potential in processing graph-structured data, their application in talent recommendation systems still faces many limitations. The computational complexity of GNN models is relatively high, especially when processing large-scale knowledge graphs, which requires a long model training time and consumes a lot of computing resources. In addition, the model design of GNNs often needs to be customized according to specific application scenarios, which increases the difficulty of development and deployment.

[0062] These problems limit the overall performance of the talent recommendation system and make it difficult to provide efficient and accurate recommendation results for financial institutions. Therefore, it is urgent to solve these key problems through technological innovation and improve the intelligence and personalization level of the talent recommendation system.

[0063] See also Figure 1 , Figure 1 This is a flowchart of an embodiment of a recommendation method based on a knowledge graph provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0064] like Figure 1 As shown, the recommendation method based on knowledge graph proposed in the present invention includes the following steps:

[0065] S10, acquiring knowledge data of the target domain, performing semantic analysis on the knowledge data through a bidirectional LSTM autoencoder, and generating a semantic vector of the knowledge data;

[0066] In this embodiment, relevant knowledge data is obtained from the target domain and necessary preprocessing is performed on it. The target domain refers to the area that needs to be focused on or applied in a specific subject or industry. The scope of this area is usually defined according to user needs or system goals. The target domain can be a collection of knowledge and information related to a specific task, problem or application. For example, the target domain may be any of the following:

[0067] Financial field: involves knowledge of financial markets, stocks, bonds, investments, corporate finance, etc.

[0068] Medical field: including patient information, medical research, diagnosis and treatment methods, drug data, etc.

[0069] Technology research and development areas: involving technological innovation, patents, academic papers, technical standards, etc.

[0070] Legal field: including legal provisions, precedents, court decisions and other information.

[0071] The knowledge data in the target domain refers to a collection of important information in a specific domain. After being collected, organized and structured, these data can be used for subsequent analysis and recommendation. For example, in a talent recommendation system, the knowledge data in the target domain may include expert information, research results, work experience, etc. related to the field. Semantic analysis of knowledge data through bidirectional LSTM autoencoder: refers to the use of natural language processing technology to deeply analyze the text data in the target domain, extract the semantic information, and represent it as a semantic vector that can be understood by the computer for subsequent graph construction and recommendation.

[0072] In this embodiment and the following embodiments, talent recommendation is used as an example for illustration. The knowledge data in the target field refers to all information and data related to a specific talent field (for example, the financial field, the technology research and development field, the medical field, etc.). Such data may include but are not limited to:

[0073] Academic literature: research papers, patent documents, technical reports, etc. related to the field. These documents contain the latest research results, technological innovations and trend analysis in the field.

[0074] Industry data: market analysis reports, industry trends, employment data, etc. related to the field. These data help to understand the market demand, development direction and talent needs in this field.

[0075] Social and professional network data: For example, data on professional social platforms such as LinkedIn, which include professional information, skills, expertise, and work experience of industry experts, scholars, and practitioners.

[0076] Educational background and training records: This includes information such as the educational background, training received, and professional certifications of people working in this field. This data helps to assess the knowledge level and professional ability of talents.

[0077] Work experience and project experience: the work experience, projects participated, roles played, achievements, etc. of talents in the target field. This information can fully reflect the actual work ability and experience accumulation of talents.

[0078] Internal company and organization data: including internal human resources data, employee performance evaluations, talent development plans, etc., which can provide in-depth insights into employee capabilities, performance and potential.

[0079] Knowledge data in the target field is a process of collecting and organizing various information related to talents in a specific field. By integrating data from the above multiple sources, a comprehensive and structured knowledge system can be constructed to reflect the skill requirements, industry trends, and the best professionals in the target field. These data play a key role in the talent recommendation system, helping the system to more accurately match and recommend outstanding talents suitable for target positions or projects.

[0080] In application scenarios, the accuracy and comprehensiveness of knowledge data directly affect the reliability and effectiveness of recommendation results. For example, in the financial field, understanding the candidate's academic background, work experience, specific skill certifications, and their influence in the industry can help companies find the financial experts or management talents that best suit their needs. Through in-depth analysis of these data, the system can better understand and predict the potential and adaptability of talents, thereby providing high-quality recommendation results.

[0081] These data may come from different sources and formats, so they need to be preprocessed to make them suitable for subsequent analysis steps. The preprocessing process includes cleaning, word segmentation, and removal of stop words to ensure the quality and consistency of the data. The purpose of this step is to provide high-quality input data for subsequent semantic analysis.

[0082] The bidirectional LSTM autoencoder is a deep learning model specifically designed to process sequence data. Through the bidirectional LSTM network, the model can process data from the beginning to the end of a sentence and from the end to the beginning of a sentence simultaneously, thereby capturing the contextual information in the text. This process generates hidden states that represent the semantics of the text, and finally generates a semantic vector through the autoencoder. The semantic vector is a compressed representation of the original text information, which contains the core semantic features of the text.

[0083] Through the processing of the bidirectional LSTM autoencoder, the original text data is converted into semantic vectors. These semantic vectors are multi-dimensional real number vectors that can compactly represent the semantic content of the text and provide support for subsequent entity recognition and relationship extraction. The semantic vectors capture important information in the text, allowing subsequent steps to perform more complex analysis and reasoning based on these vectors.

[0084] In one specific implementation, when processing text data, the model introduces an attention mechanism to dynamically assign weights to important words in the text, thereby enhancing the representation ability of the semantic vector. The attention mechanism dynamically adjusts the weight of each word according to the context of the text content, so that important words occupy a larger weight ratio in the generation of semantic vectors, thereby enhancing the semantic representation effect of the text. The key words related to the main theme of the text are strengthened through the attention mechanism, so that the generated semantic vector can better reflect the core content of the text. By introducing the attention mechanism, the semantic vector can more accurately capture the key information of the text when processing complex texts, thereby improving the semantic expression ability of the vector.

[0085] In another specific embodiment, a convolutional layer is introduced into a bidirectional LSTM autoencoder, and a convolutional neural network is used to capture local semantic features in the text, which is combined with the global semantic extraction capability of the LSTM network. The local features of the text are extracted through the convolutional layer, and the semantic features of the phrases or words in the text are captured, which are combined with the global semantic features extracted by the LSTM network. The convolutional layer performs multi-scale feature extraction on the text with different convolution kernel sizes, so that the generated semantic vector can simultaneously reflect the local and global semantic information of the text. The introduction of the convolutional layer improves the ability of the semantic vector to capture text details, especially when processing text with obvious local features, the generated semantic vector can more accurately reflect the overall semantic structure of the text.

[0086] Through the bidirectional LSTM autoencoder, the knowledge data in the target field is deeply semantically analyzed to generate high-quality semantic vectors. It can more accurately capture the semantic information in the text, providing a solid foundation for subsequent entity recognition and knowledge graph construction, thus significantly improving the accuracy and intelligence level of the talent recommendation system.

[0087] S20, combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, identifying entities and relationships between entities from the knowledge data, and constructing a triple including a subject, an object and a predicate;

[0088] In this embodiment, a bidirectional LSTM autoencoder is used to generate a semantic vector of the text, and key entities are extracted from the semantic vector using named entity recognition technology. The relationship between these entities is annotated using semantic role labeling technology, and finally a triple consisting of subject, object, and predicate is constructed. These triples serve as the basis of the knowledge graph and provide data support for subsequent talent recommendation and relationship reasoning.

[0089] The semantic vector is a multi-dimensional real number vector generated by a bidirectional LSTM autoencoder, which represents the core semantic information in the text. The generated semantic vector is used as input to provide a semantic basis for subsequent named entity recognition and semantic role labeling. These semantic vectors cover the key concepts and relationships in the text, laying a data foundation for entity recognition and relationship extraction.

[0090] Named Entity Recognition (NER) is a natural language processing technology used to identify entities mentioned in text, such as names of people, organizations, places, etc. NER technology combined with semantic vectors can accurately extract entities with specific meanings from knowledge data. These entities may represent key elements related to talent, such as people, companies, skills, etc. NER further enriches the representation capabilities of semantic vectors by classifying and annotating entity words in text.

[0091] Semantic Role Labeling (SRL) is a technique used to identify and label different roles and their semantic relationships in text. SRL combines semantic vectors with entities extracted by NER to identify semantic relationships between entities in text, such as the performer of an action, the affected person, the place, the time, etc. SRL provides the necessary semantic information for constructing triples, and accurately labels the role of each entity in a sentence by analyzing the grammatical structure and semantic relationships in the text.

[0092] A triple is a structure consisting of a subject, an object, and a predicate, which is used to represent the relationship between entities. The subject represents the executor of the action, the object represents the person affected by the action, and the predicate describes the relationship between the subject and the object. By combining NER and SRL techniques, these triples are extracted from knowledge data to construct a knowledge graph containing entities and their relationships. These triples are the basic building blocks of the knowledge graph, reflecting the various associations between entities and providing structured data for subsequent reasoning and recommendations.

[0093] In one specific implementation, a deep neural network (DNN) is introduced in the semantic role labeling process to enhance the accuracy of relationship extraction. The DNN model is used to further extract features from the semantic vector to generate a higher-dimensional representation, so that SRL can identify more complex semantic relationships. A multi-layer network structure is introduced into the DNN to capture the semantic information in the text layer by layer, so as to better identify complex entity relationships. By introducing a deep neural network, SRL can accurately identify multiple relationships between entities in complex sentence structures and improve the construction effect of triples.

[0094] In another specific implementation, in the process of constructing triples, a self-supervised learning mechanism is introduced to improve the generalization ability of the model. A large amount of unlabeled data is pre-trained using a self-supervised learning method to generate semantic vectors with more generalization ability to provide support for NER and SRL. By contrastive learning technology, the ability to distinguish between entities and relationships is enhanced, so that the model can still accurately identify and annotate entity relationships when faced with unseen data. The introduction of self-supervised learning effectively improves the adaptability of the model in new data scenarios, making the construction of triples more robust in diverse data sources.

[0095] By combining the generated semantic vectors, named entity recognition technology and semantic role labeling technology, we can efficiently and accurately identify entities and their relationships in the text and build structured triples. This improves the efficiency and accuracy of knowledge graph construction, provides a more accurate data foundation for the talent recommendation system, and significantly improves the shortcomings of existing technologies in complex relationship recognition and knowledge representation.

[0096] S30, constructing a structured knowledge graph based on the triples, extracting node features and embedding representation in the knowledge graph, and generating a feature vector for each node;

[0097] In this embodiment, based on the constructed triples, a structured graph is formed by mapping each entity and relationship to nodes and edges in the knowledge graph. In the knowledge graph, feature extraction is performed for each node to obtain feature information such as node degree and centrality. The extracted features and graph structure information are converted into low-dimensional vector form using the node embedding method to generate a feature vector for each node. The generated feature vector can be used for downstream tasks such as node classification and relationship prediction.

[0098] A triple is a structured data consisting of a subject, an object, and a predicate, and is used to represent the relationship between entities. When constructing a knowledge graph, triples are used as the basic unit of the graph. Each triple represents an edge in the graph, connecting two nodes (entities). The knowledge graph integrates multiple triples together to form a graph structure that can represent a large number of entities and their relationships. The knowledge graph constructed using triples can describe the complex relationships between entities, providing an intuitive and structured way to store and represent knowledge.

[0099] Node feature extraction refers to extracting important information related to each node in the knowledge graph, such as the node's degree, the number of connected edges, the node's centrality, etc. These features can reflect the position and importance of the node in the entire graph. For example, the node's degree reflects the connection between the node and other nodes, and the centrality measures the relative importance of the node in the entire network. By extracting these features, we can better understand the role and influence of each node in the knowledge graph.

[0100] Node embedding representation maps the nodes in the knowledge graph into a low-dimensional vector space, so that the topological information and node features in the graph structure can be expressed in the form of vectors. Through embedding representation, complex graph structure information can be converted into a vector form that is easy to calculate and analyze. These vectors not only retain the semantic and topological information in the original graph structure, but can also be directly used in subsequent machine learning models, such as node classification, link prediction and other tasks. Common node embedding methods include node2vec, DeepWalk, etc.

[0101] In one specific implementation, in the process of node embedding representation, an adaptive embedding technology is used to dynamically adjust the embedding vector of the node according to the changes in the graph structure. Through the adaptive algorithm, the changes in the graph structure are perceived, and the embedding vector of the node is dynamically adjusted, so that the representation of the node can reflect the structural changes of the graph in real time. The multi-dimensional embedding technology is introduced so that the embedding vector of the node can reflect the various characteristics and relationships of the node in different dimensions. The adaptive embedding technology enables the feature vector of the node to better adapt to the dynamically changing graph structure, improving the accuracy and flexibility of the node representation.

[0102] By combining the knowledge graph constructed by triples, feature extraction and embedding representation of nodes can be performed, which can efficiently generate vector representations that reflect graph structure information and node features.

[0103] S40, combining the feature vector of each node with the graph structure information in the knowledge graph to build an initial graph neural network model;

[0104] In this embodiment, the feature vector of each node is obtained by extracting and embedding the feature of the node, reflecting the basic attributes and local structural information of the node in the knowledge graph. The graph structure information in the knowledge graph includes the connection relationship between nodes, the degree of the node, the path information, etc. These graph structure information can not only describe the relationship between nodes, but also reveal the global topological structure of the graph. When constructing the initial graph neural network model, by combining the feature vector of each node with the graph structure information, the relative position and importance of the node in the entire knowledge graph can be captured. This combination ensures that the characteristics of the node not only rely on its own information, but also make full use of the relationship information of the surrounding neighbor nodes, thereby enhancing the comprehensiveness and accuracy of the node representation.

[0105] The initial graph neural network model takes the feature vector and graph structure information of the node as input, propagates and aggregates layer by layer, and gradually generates the global feature representation of the node. The graph neural network model can effectively perform machine learning tasks on graph structure data, propagating and aggregating node information through multi-layer networks, so that the representation of each node not only contains its own features, but also includes the features and relationship information of its neighboring nodes. In this way, the initial graph neural network model can capture complex graph structure features and generate high-quality node representations, laying the foundation for subsequent classification, prediction and other tasks.

[0106] In one specific implementation, a residual connection mechanism is introduced into the graph neural network model to improve the information transfer efficiency and stability of the model in the deep network. By adding a residual connection path to each layer of the graph neural network, the output of the previous layer is directly added to the input of the current layer to reduce information loss. The residual connection effectively alleviates the gradient vanishing problem in the deep model, ensuring that important information can be effectively retained and transmitted during the multi-layer network propagation process. The introduction of residual connections enhances the stability and information transfer efficiency of the graph neural network model, especially when dealing with deep network structures, it can better maintain the training effect and prediction performance of the model.

[0107] By combining the feature vectors of nodes with the graph structure information in the knowledge graph, an initial graph neural network model is constructed. This model effectively integrates the local features and global structure information of nodes, so that the generated node representation can better reflect its position and importance in the graph. Compared with existing graph structure processing methods, it significantly improves the representation ability and learning efficiency of graph neural networks, and can show superior performance when processing large-scale complex graphs.

[0108] S50, training and optimizing the initial graph neural network model to generate a target graph neural network model;

[0109] In this embodiment, the initial graph neural network model is trained and optimized to enable the model to better fit the training data and have good generalization capabilities in practical applications. The training process inputs the data in the training set into the model, calculates the difference between the output of the model and the true value (i.e., the loss value), and adjusts the model parameters through the back propagation algorithm to gradually reduce the loss value and finally optimize the performance of the model. The optimization process includes selecting an appropriate optimization algorithm, such as the Adam optimizer, and adjusting hyperparameters such as the learning rate, so that the model achieves the best balance between convergence speed and performance. After training and optimization, the generated target graph neural network model can better handle data tasks in practical applications.

[0110] In one specific implementation, during the training of the graph neural network model, the learning rate is dynamically adjusted to improve the convergence speed and stability of the model. Using a dynamic learning rate adjustment mechanism, during the training process, the learning rate is adaptively adjusted according to the convergence of the model, so that the model can be optimized more robustly when it is close to convergence. Use a larger learning rate to accelerate convergence in the early stage of training, and gradually reduce the learning rate when the model is close to convergence to avoid excessive step size causing model oscillation or divergence. The dynamic learning rate adjustment mechanism improves the training efficiency of the model, reduces the training time, and improves the convergence stability of the model, so that the model can reach optimal performance more quickly.

[0111] By training and optimizing the initial graph neural network model, a target graph neural network model with better performance is generated. This model can better fit the training data when processing complex graph structure data, and has good generalization ability, which is suitable for a variety of tasks in practical applications.

[0112] S60, based on the target graph neural network model, scoring and sorting the nodes representing entities in the target domain, and recommending the nodes that meet the recommendation conditions as target objects in the target domain.

[0113] In this embodiment, in the target graph neural network model, each node represents an entity in the target field, and the connection between nodes represents the relationship between entities. The scoring process calculates the importance or fitness of each node through the graph neural network model, and the basis for scoring may include factors such as the node's feature representation, centrality index, and the strength of the relationship with neighboring nodes. The sorting is based on the scoring results, and the nodes are arranged from high to low, with the highest scoring nodes being given priority as the recommended objects. In this way, the best talents who meet the recommendation conditions can be screened out from a large number of candidate nodes.

[0114] Based on the scoring and sorting, nodes with higher scores are selected as recommended objects. These nodes usually represent entities with outstanding capabilities or characteristics in a specific field. In the recommendation process, the graph neural network model combines the feature representation of the entity and the graph structure information to ensure that the recommendation results meet actual needs and have high accuracy and reliability.

[0115] In one specific implementation, a reverse verification mechanism is introduced in the node recommendation process to verify and optimize the recommendation results. After the recommendation results are generated, the rationality of the recommendation results is reversely verified by comparing the relationship between the node and its adjacent nodes, and the results that do not meet expectations are adjusted. The reverse verification results are used as feedback to iteratively update the scoring and sorting mechanism to gradually optimize the performance of the recommendation model. The reverse verification mechanism improves the reliability of the recommendation results, avoids possible deviations and errors in the model during the scoring and sorting process, and ensures that the recommended talents truly meet the needs of the target field.

[0116] By scoring and sorting nodes based on the target graph neural network model, outstanding talents in the target field can be effectively identified and screened.

[0117] The present invention relates to a recommendation method based on knowledge graph, which performs semantic analysis on knowledge data in a target field through a bidirectional LSTM autoencoder to generate a semantic vector. The semantic vector is combined with named entity recognition technology and semantic role labeling technology to identify entities and their relationships, construct triples and generate a structured knowledge graph. Node feature extraction and embedding representation are performed in the knowledge graph, an initial graph neural network model is constructed in combination with graph structure information, a target graph neural network model is generated through training and optimization, nodes are scored and sorted, and outstanding talents are recommended. The present invention can improve the accuracy and intelligence level of talent recommendation, enhance the personalization and matching degree of recommendation results, and significantly improve the problem of inaccurate recommendation results in the prior art.

[0118] In one embodiment, in the above S10, the semantic analysis of the knowledge data is performed by a bidirectional LSTM autoencoder to generate a semantic vector of the knowledge data, including:

[0119] S101, perform text segmentation and word embedding processing on the knowledge data to convert the text into a word vector with a fixed dimension;

[0120] S102, inputting the word vector into the bidirectional LSTM autoencoder, wherein the bidirectional LSTM autoencoder includes a forward LSTM network and a reverse LSTM network;

[0121] S103, processing the word vectors from the beginning to the end of the sentence in sequence through a forward LSTM network, and processing the word vectors from the end to the beginning of the sentence in sequence through a reverse LSTM network, generating a forward hidden state and a reverse hidden state respectively;

[0122] S104, at each time step, concatenating the forward hidden state and the reverse hidden state to form a bidirectional hidden state containing context information;

[0123] S105, applying an activation function to the concatenated bidirectional hidden states, and generating a global semantic vector by concatenating or pooling the processed outputs;

[0124] S106, inputting the generated global semantic vector into the decoder, reconstructing the original input sentence through the decoder, and calculating the difference between the decoder output and the original sentence as the loss value;

[0125] S107, adjusting the model parameters according to the loss value through the back propagation algorithm to optimize the semantic representation ability of the model, and finally generating a semantic vector reflecting the semantics of the knowledge data.

[0126] In this embodiment, text segmentation is to divide the original text data into independent words, in order to facilitate subsequent word embedding and semantic analysis. Word embedding processing is to convert these words into vector representations of fixed dimensions so that the words can be calculated in the model. Word embedding maps semantically similar words to similar vectors by mapping them to high-dimensional space, thereby maintaining semantic consistency in the vector space.

[0127] The bidirectional LSTM network consists of a forward LSTM network and a reverse LSTM network. The forward LSTM network processes word vectors from the beginning to the end of the sentence and generates a forward hidden state. The reverse LSTM network processes word vectors from the end to the beginning of the sentence and generates a reverse hidden state. By combining the forward and reverse hidden states, the model can capture the dependencies between contexts in the sentence.

[0128] At each time step, the forward hidden state and the reverse hidden state are concatenated to form a bidirectional hidden state containing contextual information. This concatenation method can integrate the contextual information of each word in the sentence and enhance the expressiveness of the semantic vector.

[0129] The activation function (such as ReLU, Sigmoid) is used to perform nonlinear transformation on the concatenated bidirectional hidden state, so that the semantic vector can better express complex semantic relationships. The processed output is concatenated or pooled to generate a global semantic vector, which represents the semantic information of the entire sentence.

[0130] The decoder attempts to restore the original sentence by taking the generated global semantic vector as input. The loss function calculates the difference between the decoder output and the original sentence, which is used to measure the generation quality of the model. Through the back-propagation algorithm, the model parameters are optimized according to the loss value, thereby improving the semantic representation ability of the model.

[0131] In one specific implementation, the multi-head self-attention mechanism in Transformer is used to generate word vectors, so that each word can consider the global context. A multi-head attention layer is introduced before the LSTM layer to enhance the semantic capture ability by focusing on the correlation of different parts of the sentence. The ELU activation function is applied to retain more information details in nonlinear processing, and the generated semantic vector is more expressive. The discriminator in the GAN adversarial network is used to evaluate the generated semantic vector to optimize the semantic representation quality of the generator.

[0132] This embodiment uses a bidirectional LSTM autoencoder to perform semantic analysis on knowledge data and generate high-quality semantic vectors. Compared with the traditional unidirectional model, the bidirectional LSTM can capture the context information of the sentence, improving the accuracy and meticulousness of the semantic representation. It enhances the model's ability to capture complex semantic relationships and effectively solves the shortcomings of the existing technology in semantic representation.

[0133] In one embodiment, the above S20 includes:

[0134] S201, clustering the semantic vectors representing the entity names using a hierarchical clustering algorithm based on the generated semantic vectors to obtain a plurality of name clusters;

[0135] S202, performing entity recognition on the clustered name clusters by using named entity recognition technology to identify entities representing the same actual object;

[0136] S203, clustering the semantic vectors representing the relationships between entities using a hierarchical clustering algorithm based on the generated semantic vectors to obtain a plurality of relationship clusters;

[0137] S204, performing relationship recognition on the clustered relationship clusters by using a semantic role labeling technology, and classifying relationship sentences expressing the same semantic relationship or having the same grammatical structure into a unified relationship type;

[0138] S205, constructing a triple including a subject, an object and a predicate based on the clustered entities and the unified relationship type.

[0139] In this embodiment, the semantic vector is a representation of text data converted into a vector form through semantic analysis. These vectors capture the semantic relationship between words in the text. The hierarchical clustering algorithm groups these semantic vectors so that vectors with similar semantics are aggregated together to form name clusters. Each vector in the name cluster represents a different expression that may refer to the same entity.

[0140] Named Entity Recognition (NER) technology refers to identifying entities with specific meanings from text, such as names of people, places, and organizations. After clustering, NER technology is used to further refine name clusters, identify the same actual object represented by different name expressions, and uniformly recognize them as one entity.

[0141] Similar to entity names, semantic vectors representing relations between entities can also be grouped by hierarchical clustering. Relation clusters contain vectors of relation sentences with similar semantics or structures.

[0142] Semantic role labeling (SRL) is used to identify the predicates and their related semantic roles (such as subject, object, etc.) in the sentence. In the clustered relation clusters, SRL technology is used to identify and unify relation sentences that express the same semantic relationship or have the same grammatical structure, and classify them into a unified relation type.

[0143] Finally, by combining the identified entities with the unified relationship type, a knowledge representation in the form of triples is constructed. The triples include subject (entity), object (entity) and predicate (relationship), which can clearly represent the semantic relationship in knowledge data.

[0144] This embodiment combines the generated semantic vectors, named entity recognition technology and semantic role labeling technology to accurately identify entities and their relationships from knowledge data and construct clear and rich triple-form knowledge. It significantly improves the efficiency and accuracy of knowledge graph construction, especially when processing domain-specific complex data, and can automatically and efficiently generate structured knowledge graphs.

[0145] In one embodiment, in the above S30, node feature extraction and embedding representation are performed in the knowledge graph to generate a feature vector for each node, including:

[0146] S301, extracting basic attributes of each node in the knowledge graph, wherein the basic attributes include node type, number of connected edges, node degree, and information of neighboring nodes;

[0147] S302, setting the step length and transition probability of the random walk, performing multiple random walk operations on each node through the random walk technology, and generating multiple node sequences;

[0148] S303, using the word2vec method, training the generated node sequence through a skip-gram model or a CBOW model to generate a feature vector for each node.

[0149] In this embodiment, the basic attributes of a node include the node type (such as entity type, person, organization, location, etc.), the number of connected edges (i.e., how many other nodes the node is connected to), the node degree (the connectivity of the node, which is used to measure the degree of direct connection between a node and other nodes in the graph), and the information of neighbor nodes (the characteristics of other nodes directly connected to the node). These basic attributes help to understand the position and role of the node in the knowledge graph and provide important information for subsequent feature extraction and characterization.

[0150] In an undirected graph, the degree of a node refers to the total number of edges connected to the node. Each edge connects two nodes, so if a node is directly connected to 3 other nodes, the degree of the node is 3. The degree of a node reflects the number of neighbors that the node is directly connected to in the graph. In a directed graph, the degree of a node can be divided into in-degree and out-degree. In-degree refers to how many edges point to the node, while out-degree refers to how many edges are sent from the node. The total degree is the sum of the in-degree and the out-degree. In a recommendation system, by analyzing the degree of a node, we can better understand the strength of the association between a candidate and other nodes, and thus make more accurate recommendations.

[0151] Random walk is a technique that simulates a random process and is used to traverse nodes in a graph. The step size refers to the length of a move in a random walk, while the transition probability determines the probability of going from the current node to the next node. Through multiple random walks, multiple node sequences can be generated, which reflect the local information in the graph structure and can capture the relationship between nodes.

[0152] Word2vec is a technology used to generate word vectors, which is commonly used in the field of natural language processing. The skip-gram model learns the representation of the target word by predicting the context words, while the CBOW model predicts the target word through the context words. In graph neural networks, these models are used to learn the feature vectors of nodes. Based on the generated node sequence, each node is trained through word2vec to generate a vector representation that can capture its local graph structure information.

[0153] In one specific implementation, the basic attribute extraction of nodes not only considers static information, but also dynamically changing attributes, such as the activity or timestamp of the node. This is particularly applicable to transaction networks, where the time and frequency of each transaction will affect the importance of the node. By introducing a weighted processing mechanism, the feature representation of the node can be adjusted according to these dynamic attributes, improving the model's ability to capture the importance of the node.

[0154] This embodiment extracts node features and embeds representations in the knowledge graph to generate feature vectors for each node, effectively enhancing the representation capability of nodes in the graph structure. The random walk technology combined with the word2vec model can capture local and global information of nodes in the graph, thereby generating more expressive feature vectors. It is particularly suitable for processing complex financial knowledge graphs and can significantly improve the accuracy and efficiency of graph structure analysis.

[0155] In one embodiment, the above S40 includes:

[0156] S401, multiplying the feature vector of each node by the pre-trained weight matrix to generate a preliminary score value of the node, and adjusting the preliminary score value according to the graph structure information in the knowledge graph;

[0157] S402, calculating the weight value of each node, and performing nonlinear processing through an activation function to obtain an activated weight value;

[0158] S403, normalizing the activated weight values ​​so that the sum of the weight values ​​of each node is 1, and generating a normalized weight vector;

[0159] S404, using the normalized weight vector and the in-degree of the node, weighted adjustment is performed on the preliminary score value of the node to generate a final feature representation of the node;

[0160] S405, in each layer of the graph neural network, the final feature representation of the node and the normalized weight vector are used to control the intensity of information propagation, and the information is aggregated and updated through the network layer to finally build a complete initial graph neural network model.

[0161] In this embodiment, the feature vector of each node is generated by the previous node embedding representation, which represents the attributes of the node in the knowledge graph and its relationship with other nodes. The pre-trained weight matrix contains the initial model parameters, which is used to convert the feature vector into a score value. The preliminary score value obtained by multiplying the feature vector with the weight matrix represents the preliminary importance or priority of the node. Subsequently, based on the graph structure information of the knowledge graph, such as the neighbor relationship of the node and the number of connected edges, the preliminary score value is adjusted to more accurately reflect the status and influence of the node in the entire graph.

[0162] The calculation of the weight value is based on the preliminary score value of the node and other related information (such as the characteristics of the node, the number of connected edges, etc.). The role of the activation function is to introduce nonlinearity so that the distribution of the weight value is more uniform or consistent with the expected distribution. This nonlinear processing step helps to improve the expressiveness of the model, enabling it to handle more complex relationships and dependencies between nodes.

[0163] The purpose of normalization is to ensure that the sum of the weights of all nodes is equal to 1, so that the influence of the nodes can be fairly distributed throughout the graph structure. The generated normalized weight vector is used for subsequent weighted calculations, so that the final feature representation of each node can reflect its actual importance in the graph.

[0164] In-degree refers to the number of other nodes that a node is connected to in the graph. By considering the in-degree, the status and influence of the node can be more accurately reflected. The final feature representation is the result of further weighted adjustment of the preliminary score value, combining the normalized weight vector and in-degree information. This feature representation will be used to construct each layer of the graph neural network.

[0165] In each layer of the graph neural network, the final feature representation of the node and the normalized weight vector jointly determine the propagation path and strength of the information. The network layer is responsible for aggregating and updating the information received from neighboring nodes, transmitting and processing the information between nodes layer by layer, and finally forming a complete graph neural network model. This model can efficiently capture and express complex relationships in the graph structure, providing support for subsequent tasks such as prediction and classification.

[0166] In one specific implementation, the weight matrix is ​​no longer a static pre-trained matrix, but is dynamically updated according to the model feedback of each training round. By updating the weight matrix after each round of iteration, the complex relationship between nodes can be more accurately reflected, making the calculation of the preliminary score value closer to the actual situation.

[0167] This embodiment combines the feature vector of each node with the graph structure information in the knowledge graph to construct an initial graph neural network model, which can efficiently capture the complex relationships in the graph structure and generate accurate node feature representations. The model can gradually optimize node representation and model performance through multi-layer network propagation and information aggregation, and finally form an efficient and accurate graph neural network model. The performance of the model in prediction, classification and other tasks is significantly improved.

[0168] In one embodiment, the above S50 includes:

[0169] S501, using mean square error as the loss function, calculating the prediction error of the initial graph neural network model on the training set;

[0170] S502, using the calculated prediction error, adjusting the model parameters using the Adam optimizer through the back propagation algorithm to minimize the loss function;

[0171] S503, performing multiple iterations of training, and after each training round, using the validation set to evaluate and adjust the model performance;

[0172] S504, after multiple rounds of training and optimization, an optimized target graph neural network model is finally generated for prediction and recommendation tasks.

[0173] In this embodiment, mean square error (MSE) is a commonly used loss function used to measure the difference between the model prediction value and the actual value. Specifically, MSE evaluates the performance of the model by calculating the square average of the difference between the predicted value and the actual value. The smaller the mean square error, the closer the model's prediction is to the actual value. By using MSE as a loss function, the parameter adjustment of the model can be effectively guided to make the prediction result more accurate.

[0174] The back propagation algorithm is a key algorithm for training neural networks. It is used to calculate the gradient of the loss function and update the model parameters based on the gradient information. The Adam optimizer is an optimization algorithm based on adaptive learning rate. It can combine the momentum method and the RMSProp method to converge to the optimal solution quickly and stably. By combining the Adam optimizer and the back propagation algorithm, the model can continuously adjust the parameters and gradually reduce the value of the loss function, thereby improving the model's predictive ability.

[0175] Iterative training is a common method in neural network training. Through multiple rounds of training, the model gradually adjusts parameters and optimizes performance. After each round, the validation set (i.e., the data set that is not involved in the training) is used to evaluate the performance of the model to ensure that the model is not overfitted and can perform well in practical applications. Through this process, the model can gradually approach the optimal state after multiple rounds of training.

[0176] After multiple rounds of training and optimization, the model finally converged to a stable state, that is, the optimal graph neural network model. This optimized model has strong generalization ability and can be effectively applied to actual prediction and talent recommendation tasks. The model provides support for talent recommendation by capturing potential patterns and relationships in the data.

[0177] This embodiment trains and optimizes the initial graph neural network model to generate a target graph neural network model with high prediction accuracy and generalization ability. It can effectively capture complex patterns in the data and adapt to different data distributions, thus performing well in actual prediction and talent recommendation tasks, and significantly improving the stability and accuracy of the model.

[0178] In one embodiment, in the above S60, scoring and sorting nodes representing entities in the target domain based on the target graph neural network model includes:

[0179] S601, extracting a final feature representation of each node from the target graph neural network model, wherein the final feature representation is a vector generated through a multi-layer network propagation and information aggregation process;

[0180] S602, calculating the centrality index of the relative importance of each node in the graph structure by using degree centrality, closeness centrality, betweenness centrality or eigenvector centrality algorithm;

[0181] S603, using a preset transformation function, combining the feature vector of each node with the centrality index to generate a preliminary evaluation value of each node;

[0182] S604, analyzing the connection strength between the node and its neighboring nodes, and adjusting the preliminary evaluation value of each node according to the analysis result to generate a final evaluation value;

[0183] S605 , sorting the nodes representing entities in the target domain based on the final evaluation value of each node.

[0184] In this embodiment, the final feature representation of the node is a vector generated by the multi-layer network propagation and information aggregation process in the target graph neural network model. These feature representations are a comprehensive description of the node in the graph structure, including the attribute information of the node and its relationship with other nodes. These vectors represent the position of the node in the graph and its connectivity and importance with other nodes.

[0185] Centrality metrics are a class of algorithms used to measure the importance or influence of a node in the entire graph structure. Degree centrality measures how many other nodes a node is connected to; closeness centrality evaluates the shortest path distance from a node to all other nodes; betweenness centrality measures the criticality of a node in the graph by calculating how many shortest paths a node is on; and eigenvector centrality considers the importance of a node and its neighboring nodes. Through these centrality algorithms, an indicator can be generated for each node that reflects its position in the graph structure.

[0186] Transformation functions are used to combine the feature vector and centrality index of a node to generate a preliminary evaluation value for the node. These functions can be linear or nonlinear functions, designed according to the requirements of a specific task. The preliminary evaluation value represents the relative importance or ranking of the node under the current graph neural network model.

[0187] The connection strength reflects the strength of the relationship between a node and its neighboring nodes. This may include edge weights, relationship types, etc. After analyzing these connection strengths, the preliminary evaluation values ​​are adjusted so that the final evaluation values ​​more accurately reflect the node's position and influence in the entire graph.

[0188] The final evaluation value is used to sort the nodes. This sorting process arranges the nodes according to their relative importance or influence based on their final evaluation value, thereby determining which nodes have higher priority or recommendation value in the target field.

[0189] In another specific implementation, when analyzing the connection strength between a node and its neighboring nodes, not only the weight of the edge is considered, but also the time dimension is added to analyze the persistence and frequency of the connection. This strategy can better identify the strength of the relationship between nodes, thereby making more accurate adjustments to the preliminary evaluation value. When adjusting the preliminary evaluation value of a node, the final evaluation value of the node is further optimized in combination with external expert data or domain knowledge, so that the results of the model are more in line with the actual needs of the industry.

[0190] This embodiment can effectively identify and recommend key talent nodes in the field by scoring and ranking the nodes representing entities in the target field based on the target graph neural network model. Not only does it comprehensively consider the feature representation and centrality index of the nodes, but it also further optimizes the ranking of the nodes through the analysis of the connection strength. This multi-level and multi-dimensional analysis method makes talent recommendations more accurate and helps improve the accuracy and efficiency of decision-making.

[0191] The present invention also provides a recommendation device based on knowledge graph, referring to Figure 2 , Figure 2 This is a functional module diagram of a preferred embodiment of a recommendation device based on a knowledge graph of the present invention. The recommendation device based on a knowledge graph includes:

[0192] The data processing and semantic analysis module acquires the knowledge data of the target domain, performs semantic analysis on the knowledge data through a bidirectional LSTM autoencoder, and generates a semantic vector of the knowledge data;

[0193] An entity recognition and relationship extraction module, combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, identifies entities and relationships between entities from the knowledge data, and constructs a triple containing a subject, an object and a predicate;

[0194] A knowledge graph construction module constructs a structured knowledge graph based on the triples, extracts node features and embeds representations in the knowledge graph, and generates a feature vector for each node;

[0195] The graph neural network construction module combines the feature vector of each node with the graph structure information in the knowledge graph to build the initial graph neural network model;

[0196] A graph neural network training module, which trains and optimizes the initial graph neural network model to generate a target graph neural network model;

[0197] The target recommendation module scores and sorts the nodes representing entities in the target domain based on the target graph neural network model, and recommends the nodes that meet the recommendation conditions as target objects in the target domain.

[0198] The specific implementation of the knowledge graph-based recommendation device of the present invention is basically the same as the above-mentioned embodiments of the knowledge graph-based recommendation method, and will not be repeated here.

[0199] The present invention also provides a computer device, such as Figure 3 As shown, the computer device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0200] Those skilled in the art will understand that Figure 3 The hardware structure of the computer device shown in the figure does not constitute a limitation on the computer device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0201] like Figure 3 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a recommendation program based on a knowledge graph. The operating system is a program for managing and controlling computer devices and software resources, and supports the operation of the network communication module, the user interface module, the recommendation program based on the knowledge graph, and other programs or software; the network communication module is used to manage and control the network interface 1004; and the user interface module is used to manage and control the user interface 1003.

[0202] exist Figure 3In the computer device hardware structure shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the client and communicate data with the client; the processor 1001 can call the knowledge graph-based recommendation program stored in the memory 1005, and perform the same operation as the knowledge graph-based recommendation method.

[0203] The specific implementation of the computer device of the present invention is basically the same as the above-mentioned embodiments of the recommendation method based on the knowledge graph, and will not be repeated here.

[0204] In addition, an embodiment of the present invention also proposes a computer storage medium, on which a recommendation program based on a knowledge graph is stored. When the recommendation program based on the knowledge graph is executed by a processor, the steps of the recommendation method based on the knowledge graph as described above are implemented.

[0205] The specific implementation methods of the computer storage medium of the present invention are basically the same as the above-mentioned embodiments of the recommendation method based on the knowledge graph, and will not be repeated here.

Claims

1. A recommendation method based on knowledge graph, characterized in that: The following steps are involved: Acquire knowledge data in a target domain, perform semantic analysis on the knowledge data through a bidirectional LSTM autoencoder, and generate a semantic vector of the knowledge data; Combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, identifying entities and relationships between entities from the knowledge data, and constructing a triple containing a subject, an object and a predicate; Building a structured knowledge graph based on the triples, extracting node features and embedding representation in the knowledge graph, and generating a feature vector for each node; Combine the feature vector of each node with the graph structure information in the knowledge graph to build the initial graph neural network model; Training and optimizing the initial graph neural network model to generate a target graph neural network model; Based on the target graph neural network model, the nodes representing entities in the target domain are scored and sorted, and the nodes that meet the recommendation conditions are recommended as target objects in the target domain.

2. The recommendation method based on knowledge graph according to claim 1, characterized in that: Performing semantic analysis on the knowledge data through a bidirectional LSTM autoencoder to generate a semantic vector of the knowledge data includes: Perform text segmentation and word embedding processing on knowledge data to convert text into word vectors of fixed dimensions; Inputting the word vector into the bidirectional LSTM autoencoder, wherein the bidirectional LSTM autoencoder includes a forward LSTM network and a reverse LSTM network; The word vectors are processed from the beginning to the end of the sentence through the forward LSTM network, and the word vectors are processed from the end to the beginning of the sentence through the reverse LSTM network to generate the forward hidden state and the reverse hidden state respectively; At each time step, the forward hidden state and the reverse hidden state are concatenated to form a bidirectional hidden state that contains contextual information. Apply activation function to the concatenated bidirectional hidden state, and generate a global semantic vector through concatenation or pooling of the processed output; The generated global semantic vector is input into the decoder, and the original input sentence is reconstructed through the decoder, and the difference between the decoder output and the original sentence is calculated as the loss value; The model parameters are adjusted according to the loss value through the back-propagation algorithm to optimize the semantic representation ability of the model, and finally a semantic vector reflecting the semantics of the knowledge data is generated.

3. The recommendation method based on knowledge graph according to claim 1, characterized in that: Combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, entities and relationships between entities are identified from the knowledge data, and triples containing subject, object and predicate are constructed, including: Based on the generated semantic vectors, a hierarchical clustering algorithm is used to cluster the semantic vectors representing entity names to obtain multiple name clusters; The named entity recognition technology is used to perform entity recognition on the clustered name clusters to identify entities representing the same actual objects; Based on the generated semantic vectors, a hierarchical clustering algorithm is used to cluster the semantic vectors representing the relationship between entities to obtain multiple relationship clusters; The semantic role labeling technology is used to identify the relationship clusters after clustering, and the relationship sentences expressing the same semantic relationship or having the same grammatical structure are classified into a unified relationship type; Based on the clustered entities and the unified relationship type, a triple consisting of subject, object and predicate is constructed.

4. The recommendation method based on knowledge graph according to claim 1, characterized in that: Node feature extraction and embedding representation are performed in the knowledge graph to generate a feature vector for each node, including: Extracting basic attributes of each node in the knowledge graph, wherein the basic attributes include node type, number of connected edges, node degree, and information of neighboring nodes; Set the step size and transfer probability of the random walk, and use the random walk technology to perform multiple random walk operations on each node to generate multiple node sequences; Using the word2vec method, the generated node sequence is trained through the skip-gram model or the CBOW model to generate the feature vector of each node.

5. The knowledge graph-based recommendation method as claimed in claim 1, characterized in that: Combining the feature vector of each node with the graph structure information in the knowledge graph, we build the initial graph neural network model, including: Multiply the feature vector of each node by the pre-trained weight matrix to generate a preliminary score value for the node, and adjust the preliminary score value based on the graph structure information in the knowledge graph; Calculate the weight value of each node, and perform nonlinear processing through the activation function to obtain the weight value after activation; Normalize the activated weight values ​​so that the sum of the weight values ​​of each node is 1, and generate a normalized weight vector; Using the normalized weight vector and the node's in-degree, the node's preliminary score value is weighted and adjusted to generate the node's final feature representation; In each layer of the graph neural network, the final feature representation of the node and the normalized weight vector are used to control the intensity of information propagation, and information is aggregated and updated through the network layer to finally build a complete initial graph neural network model.

6. The recommendation method based on knowledge graph according to claim 1, characterized in that: The initial graph neural network model is trained and optimized to generate a target graph neural network model, including: Use mean square error as the loss function to calculate the prediction error of the initial graph neural network model on the training set; Using the calculated prediction error, the model parameters are adjusted using the Adam optimizer through the back-propagation algorithm to minimize the loss function. Perform multiple iterations of training, and use the validation set to evaluate and adjust the model performance after each training round. After multiple rounds of training and optimization, the optimized target graph neural network model is finally generated for prediction and recommendation tasks.

7. The recommendation method based on knowledge graph according to claim 1, characterized in that: Based on the target graph neural network model, nodes representing entities in the target domain are scored and sorted, including: Extracting a final feature representation of each node from the target graph neural network model, wherein the final feature representation is a vector generated through a multi-layer network propagation and information aggregation process; Calculate the centrality index of the relative importance of each node in the graph structure through degree centrality, closeness centrality, betweenness centrality or eigenvector centrality algorithms; Using the preset transformation function, the feature vector of each node is combined with the centrality index to generate a preliminary evaluation value for each node; Analyze the connection strength between the node and its neighboring nodes, and adjust the preliminary evaluation value of each node based on the analysis results to generate the final evaluation value; The nodes representing entities in the target domain are ranked based on the final evaluation value of each node.

8. A recommendation device based on knowledge graph, characterized in that: The recommendation device based on knowledge graph includes: The data processing and semantic analysis module acquires the knowledge data of the target domain, performs semantic analysis on the knowledge data through a bidirectional LSTM autoencoder, and generates a semantic vector of the knowledge data; An entity recognition and relationship extraction module, combining the generated semantic vector, named entity recognition technology and semantic role labeling technology, identifies entities and relationships between entities from the knowledge data, and constructs a triple containing a subject, an object and a predicate; A knowledge graph construction module constructs a structured knowledge graph based on the triples, extracts node features and embeds representations in the knowledge graph, and generates a feature vector for each node; The graph neural network construction module combines the feature vector of each node with the graph structure information in the knowledge graph to build the initial graph neural network model; A graph neural network training module, which trains and optimizes the initial graph neural network model to generate a target graph neural network model; The target recommendation module scores and sorts the nodes representing entities in the target domain based on the target graph neural network model, and recommends the nodes that meet the recommendation conditions as target objects in the target domain.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a knowledge graph-based recommendation program stored in the memory and executable on the processor. When the knowledge graph-based recommendation program is executed by the processor, the steps of the knowledge graph-based recommendation method as described in any one of claims 1 to 7 are implemented.

10. A computer storage medium, characterized in that: The storage medium stores a recommendation program based on a knowledge graph, and when the recommendation program based on a knowledge graph is executed by a processor, the steps of the recommendation method based on a knowledge graph as described in any one of claims 1 to 7 are implemented.

Images