A hierarchical approach to large-scale complex semantic networks

Abstract

A hierarchical method for large-scale complex semantic networks, including four steps: network structure semantic information modeling, multi-view clustering to extract mixed cohesive layers, structural cohesive layer extraction, and mixed cohesive node summary extraction. The present invention proposes a network hierarchical method, which is a cohesive node extraction method that integrates semantic and structural features, constructs three levels of different scales for the network, and designs a hierarchical visualization method that supports large-scale complex semantic network exploration. In terms of technical means, a graph representation learning method is used to extract structural and semantic embedding vectors, and regards them as different view features of the network. On this basis, the present invention designs a multi-view clustering algorithm to extract cohesive node clusters, and further aggregates them into mixed cohesive nodes. Multiple cohesive nodes are linked to each other to form a coarse-grained hierarchy.

Description

Technical Field

[0001] The invention relates to the technical field of data exploration, and in particular to a hierarchical method for large-scale complex semantic networks. Background Art

[0002] The rapid development of information technology has also brought about tremendous changes in the knowledge structure. From the successful sending of the first email, to the rise of instant messaging tools, to the promotion of online networks, human life has broken free from the shackles of geography and extended from the real world to the virtual network. With information technology as a medium, knowledge exchange has gradually shown the characteristics of wide range, fast speed and large amount of information. Widely disseminated knowledge has formed a large-scale network with rich text information in the nodes, that is, a large-scale complex semantic network.

[0003] Faced with such a huge amount of data, the limited perception ability of humans has brought bottlenecks to their knowledge reception, so it is meaningful to find a suitable way to assist exploration activities. Visualization relies on technical means such as data analysis, computer graphics, and human-computer interaction to assist users in quickly mining information from complex networks, so that human cognitive behavior is no longer limited by the perception system, greatly improving the ability to receive information.

[0004] However, there are challenges in accurately encoding large-scale complex semantic network data into easily perceived visual elements. Large-scale complex semantic networks have two prominent features: huge network scale and complex node content. They are both difficulties that need to be solved in visual representation: affected by the scale of the network, the number of visual presentation elements often exceeds the user's perceptual ability to understand the ontological visual representation. When information overload occurs, the quality of user decision-making is likely to decline; the complexity of node semantics adds new challenges to network visualization. The rich text information carried by the nodes themselves and the complex semantic relationships in the network context are difficult to display in a limited interface, and it is extremely difficult to locate information in a huge network. Therefore, in order to meet the needs of exploring and analyzing large-scale complex semantic networks, it is necessary to find a suitable visualization form that can simplify network representation and reveal the patterns hidden behind the linked data.

[0005] As a common method to abstract and simplify networks, network hierarchical visualization regards the relationship between entities in node links as several interdependent subsystems (layers), abstracts the underlying graph structure and generates condensed nodes (also called super nodes or meta nodes in some works), and displays information at different granularities. It allows users to focus on specific areas in network visualization and explore their internal details by controlling the expansion and contraction of nodes in different levels, meeting the exploration needs of large-scale networks. However, most existing hierarchical technologies are oriented to homogeneous networks, treating network nodes as uniquely identified vertices in the topology; or oriented to attribute networks, node aggregation is performed according to the numerical attributes of the network. As an important means to simplify network representation in the field of visualization, graph hierarchical technology has made certain contributions to expanding the visual scalability boundary of the network, which can effectively save layout computing resources and improve visual quality. The emergence of large-scale networks carrying text information has put forward new requirements and challenges for their ability to process rich information and cope with large-scale data. How to scientifically extract hierarchical condensed nodes in networks rich in text information and convey their overview information to users as much as possible when displaying them is an important problem to be solved at present. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a large-scale complex semantic network hierarchical visualization method based on graph representation learning and multi-view visualization.

[0007] The present invention solves the technical problem by adopting the following technical solutions:

[0008] A hierarchical method for large-scale complex semantic networks, comprising the following steps:

[0009] S1. Network structure semantic information modeling:

[0010] Extract node embedding vectors from the network to model the network structure and semantic information, preprocess the semantic information data of the network, and use the context-dependent network representation learning model CANE to generate a low-dimensional embedding representation with network node structure and semantic information to generate feature vectors for downstream tasks;

[0011] S2. Multi-view clustering to extract mixed cohesive layers:

[0012] Multi-view clustering is used to further construct node clusters corresponding to hybrid cohesive nodes, namely, hybrid cohesive layers, with the help of feature vectors obtained from upstream tasks;

[0013] S3, structural coacervate extraction:

[0014] Using structural information to further divide node clusters makes the newly divided sub-cluster structure tend to be dispersed, that is, when the overall number of clusters increases to a certain extent, the node modularity will drop sharply, which also means that most nodes become discrete small structures and lose the connection information between them. Considering that the clusters obtained in the upper layer have semantic similarities, the further subdivision of the content under the same semantic theme can be ignored, and only the structural features can be referred to. The traditional community structure is used to further divide the network. Compared with the clustering algorithm, it reveals the dense organization inside the network and tends to aggregate nodes rather than separate them. The result obtained is called the structural cohesive layer;

[0015] S4. Hybrid Agglomerative Node Summary Extraction:

[0016] For the mixed cohesive layer, each node represents a semantic structure cluster composed of multiple sub-nodes. In order to help users get a global overview of the data without expanding the cohesive nodes, and expand the cohesive nodes according to their needs, so as to guide users to mine the network with an "overview + details" exploration behavior mode, each mixed cohesive node represents a semantic structure cluster, and a three-part information overview is generated from the inside to the outside around this cluster. For the inside of the cluster, the summary shows the user the theme of the local network represented by the cohesive node, and extracts the representative sub-nodes with important influence as special attention objects. For the outside of the cluster, the neighborhood level is displayed to reflect its closeness to other node topics in the global network, which is conducive to exploring the correlation between different topics.

[0017] Preferably, the embedding object of the network representation learning model CANE carries nodes in a complex text network and generates structural and semantic feature vectors with consistent expression forms. The node pairs linked by each edge are taken as input. The generated node semantic representation introduces a mutual attention mechanism according to the different node interaction neighbors, and has context-related features.

[0018] Preferably, the multi-view clustering method reveals the intrinsic connection and complementary information between feature vectors obtained from upstream tasks from different perspectives, and seeks complementarity and consensus between views to assist clustering.

[0019] Preferably, in the network structure semantic information modeling step, the data preprocessing method is:

[0020] The node text is case-standardized, grammatically eliminated, and stop words are removed. Stop words include functional words such as articles, prepositions, pronouns, and conjunctions.

[0021] Preferably, the multi-view clustering is implemented by the neighbor propagation algorithm, which is an unsupervised machine learning algorithm. Compared with other traditional clustering methods, the neighbor propagation algorithm does not need to specify the number of clusters and avoids the selection of k initial cluster centers. Clusters are created by sending and propagating messages between data points until convergence.

[0022] The accuracy of the similarity matrix S determines the quality of the AP algorithm execution. If the similarity matrix can well define the similarity between nodes, a good clustering result will be obtained. In the multi-view AP clustering process, due to the different physical meanings of features in different views, the similarity measurement between nodes is somewhat difficult. Before data clustering, the "consensus" between multi-view features is found, and the consensus features are used to generate a similarity matrix for clustering. The feature fusion method is used to extract common hidden features from different views, and it is used as the upstream method of AP clustering to build a more accurate similarity matrix for it.

[0023] Canonical Correlation Analysis (CCA) is a typical feature fusion method that can find paired projections for different views to maximize the correlation between them. In the dual-view analysis of words and link structures from documents, the hierarchical clustering algorithm based on CCA performs better on such data than the hierarchical clustering algorithm based on traditional PCA. Therefore, CCA is used to find typical vectors from different views to the common space.

[0024] Preferably, in the structural cohesive layer extraction step, the community structure is further divided into networks by a non-overlapping node community discovery algorithm.

[0025] The advantages and positive effects of the present invention are:

[0026] 1. The present invention proposes a network hierarchical method, which is a condensed node extraction method based on semantic and structural features, and constructs three levels of different scales for the network. In a large-scale semantic network, the hierarchical method not only reduces the visual complexity of the view layout by node contraction, but also keeps the continuation of the substructure content as much as possible while simplifying the structure. The present invention extracts a summary of the node that condenses its structural and semantic information, filters out most of the information, and retains some prominent features of the condensed sub-nodes.

[0027] 2. The present invention designs a hierarchical visualization method that supports large-scale complex semantic network exploration. In terms of technical means, a graph representation learning method is used to extract structural and semantic embedding vectors, and they are regarded as different view features of the network. On this basis, the present invention designs a multi-view clustering algorithm to extract clusters of condensed nodes, and further aggregates them into mixed condensed nodes. Multiple condensed nodes are linked to each other to form a coarse-grained hierarchy, which not only considers the structural topology, but also should be analyzed in combination with the semantics of the text content contained in the node. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of multi-view clustering of the present invention;

[0029] Figure 2 is a data analysis flow chart of the network layering method of the present invention;

[0030] Figure 3 It is a schematic diagram of the process of converting different view features in the network into consensus features in the present invention;

[0031] Figure 4 is a schematic diagram of a mixed coacervate of the present invention;

[0032] Figure 5 is a schematic diagram of a structured coacervate of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] It should be noted that when a component is referred to as being "fixed to" another component, it may be directly on the other component or there may also be a component centered. When a component is considered to be "connected to" another component, it may be directly connected to the other component or there may also be a component centered. When a component is considered to be "set on" another component, it may be directly set on the other component or there may also be a component centered. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0036] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0037] The hierarchical method for large-scale complex semantic networks described in the present invention aims to find an effective way to express the macro structure of complex semantic networks. This type of network has two characteristics: "complex node content" and "large network scale". Among them, "complex node content" specifically refers to the network nodes for which the present invention is directed to carry a large amount of text information, and "large network scale" means that the network objects used in the present invention have thousands or even tens of thousands of nodes, and such a large amount of information often exceeds the scope that users can receive, process and effectively use, so this hierarchical method is proposed.

[0038] This paper uses the Cora dataset, which is a citation network containing 2708 machine learning papers and 5429 links. The papers mainly involve machine learning subfields and were published before 2000. The overall experimental process is as follows: Figure 2 As shown, the hierarchical method of a large-scale complex semantic network described in the present invention comprises the following steps:

[0039] The goal of this invention is network stratification, which is achieved by shrinking and aggregating nodes, so the focus is on node embedding. Node embedding means that each vertex in the graph is represented by a low-dimensional latent feature vector. In the embedding space, vertices that are close to each other are more similar, and vertices that are far away from each other have greater differences. For a graph G = (v, e), each node embedding vector is a mapping of the corresponding vertex in the graph:

[0040]

[0041] Among them, v i represents the i-th vertex, y i Represents the embedding vector corresponding to i vertices. Embedding captures graph topology, relationships between nodes, and deeper related information. How to capture the inherent information of these graphs depends on the optimization goal proposed for the neural network during training.

[0042] S1. Network structure semantic information modeling:

[0043] Node embedding vectors are extracted from the network to model the network structure and semantic information, and the semantic information data of the network is preprocessed, that is, the node text is case-standardized, grammatical elimination, and stop word removal operations are performed. Stop words include functional words such as articles, prepositions, pronouns, and conjunctions. Data preprocessing is an important part of improving the quality of embedding vectors. It can help filter out a large amount of redundant information and enable downstream models to focus on important text information processing. The network representation learning model CANE related to the following text is used to generate a low-dimensional embedding representation with network node structure and semantic information to generate feature vectors for downstream tasks.

[0044] The embedding object of the network representation learning model CANE is a node in a complex text network, and generates a structural and semantic feature vector with consistent expression. The node pair linked by each edge is taken as input. The generated node semantic representation introduces a mutual attention mechanism according to the different node interaction neighbors, and has context-related features.

[0045] S2. Multi-view clustering to extract mixed cohesive layers:

[0046] Multi-view Clustering (MVC) is used to further construct node clusters corresponding to hybrid cohesive nodes with the help of feature vectors obtained from upstream tasks, namely, hybrid cohesive layers. The multi-view clustering method reveals the intrinsic connections and complementary information between feature vectors obtained from upstream tasks from different perspectives, and seeks complementarity and consensus between views to assist clustering.

[0047] like Figure 1 In the embodiment of the present invention shown, the structural view and semantic view of a network are mapped to the potential data space, part A and part C constitute the complementary parts of the views, part B is the consensus of the views, and multi-view clustering can maximize the consistency of views between different views (part B) and reach a broad consensus when forming clusters from multiple perspectives.

[0048] Multi-view clustering is implemented through the neighbor propagation algorithm, which is an unsupervised machine learning algorithm. Compared with other traditional clustering methods, the neighbor propagation algorithm does not need to specify the number of clusters and avoids the selection of k initial cluster centers. It creates clusters by sending and propagating messages between data points until convergence.

[0049] The accuracy of the similarity matrix S determines the quality of the AP algorithm execution. If the similarity matrix can well define the similarity between nodes, a good clustering result will be obtained. In the multi-view AP clustering process, due to the different physical meanings of features in different views, the similarity measurement between nodes is somewhat difficult. Before data clustering, the "consensus" between multi-view features is found, and the consensus features are used to generate a similarity matrix for clustering. The feature fusion method is used to extract common hidden features from different views, and it is used as the upstream method of AP clustering to build a more accurate similarity matrix for it.

[0050] Canonical Correlation Analysis (CCA) is a typical feature fusion method that can find paired projections for different views to maximize the correlation between them. In the dual-view analysis of words and link structures from documents, the hierarchical clustering algorithm based on CCA performs better on such data than the hierarchical clustering algorithm based on traditional PCA. Therefore, CCA is used to find typical vectors from different views to the common space, such as Figure 3 As shown in Figure 1, CCA learns the projection direction of a pair of related features by maximizing the correlation between structural and semantic features. The feature extraction result can be obtained by projecting the features using the learned projection direction.

[0051] S3, structural coacervate extraction:

[0052] Although the mixed cohesive layer obtained by steps S1 and S2 is greatly simplified compared with the original data, each cohesive node has hundreds or thousands of child nodes. Such a large amount of information will cause a confusing visual interface, difficult to understand information, and important information to be hidden. It is not enough to perform node cohesion only once, and a new layer needs to be added to simplify the data subset.

[0053] The node clusters are further divided using semantic and structural information, so that the structure of the newly divided sub-clusters tends to be dispersed. That is, when the overall number of clusters increases to a certain level, the node modularity will drop sharply, which also means that most nodes become discrete small structures and lose the connection information between them. Considering that the clusters obtained in the upper layer themselves have semantic similarities, the further subdivision of the content under the same semantic theme can be ignored, and only the structural features need to be referred to. The traditional community structure is used to further divide the network, mainly because compared with the clustering algorithm, it reveals the dense organization within the network and tends to aggregate rather than separate nodes. The result obtained is called the structural cohesive layer.

[0054] In this embodiment, the community generation algorithm selects the LPA algorithm because the time complexity of LPA is linear, which is sufficient to adapt to the heavy task of partitioning multiple cohesive nodes, and can adaptively generate community groups of different sizes according to data characteristics without pre-defining the objective function and the number of communities as prior information.

[0055] S4. Hybrid Agglomerative Node Summary Extraction:

[0056] For the mixed cohesive layer, each node represents a semantic structure cluster composed of multiple sub-nodes. In order to help users get a global overview of the data without expanding the cohesive nodes, and expand the cohesive nodes according to their needs, so as to guide users to mine the network with an "overview + details" exploration behavior mode, each mixed cohesive node represents a semantic structure cluster, and a three-part information overview is generated from the inside to the outside around this cluster. For the inside of the cluster, the summary shows the user the theme of the local network represented by the cohesive node, and extracts the representative sub-nodes with important influence as special attention objects. For the outside of the cluster, the neighborhood level is displayed to reflect its closeness to other node topics in the global network, which is conducive to exploring the correlation between different topics.

[0057] In order to scientifically extract influential nodes in the network, the present invention proposes a node importance scoring method that combines local information and global information. Faced with large-scale complex semantic networks, it is necessary to formulate a node importance evaluation method in combination with its data characteristics. The large scale of the network means that the calculation efficiency of the global importance score should be as high as possible, and it is best to get the scores of all nodes at one time, without calculating each node one by one. The method based on k-shell decomposition has a linear time complexity of O(m), and only needs to traverse all edges in the network once to obtain the importance scores of all nodes. The present invention optimizes and evaluates the global importance of nodes based on this algorithm. Taking advantage of the fact that the network carries a large amount of text information, a new metric is proposed to characterize the local importance of nodes. Different from the traditional method based on local topology and information dissemination capabilities, the method regards each node as an object with information sharing capabilities. By evaluating how much information it can bring to other neighborhood nodes, its information quality and importance can be judged.

[0058] Through visualization, the Cora dataset can be obtained after this data analysis process. Figure 4 The mixed coacervate and Figure 5 The structure of the coacervate is shown. Figure 4It can be found from the mixed cohesive layer that the data scale of different cohesive points shows that in the corpus, cluster A has the largest number of papers, with 401 papers, mainly involving networks, data, training and time; the mixed cohesive point E with the keywords "error, data, distribution, category, theory" contains the least number of papers, with only 169 papers. This shows that work in different sub-research fields in the field of machine learning cites each other, but the research methods have different focuses. Before 2000, research on network model training had already occupied the largest proportion in the field of machine learning. On the other hand, other sub-fields have a certain number of research results, but there is still a big gap compared with group A. Expand the mixed cohesive points as needed in the mixed zoom view to obtain the network composition of the structural cohesive layer and the mixed cohesive layer. Figure 5 The hybrid view shown in the figure, further selecting the structural cohesive node and expanding it, will obtain the expanded view of part of the actual network. Looking at the corresponding information about the structural cohesive node, it can be found that the research content has a strong correlation and is closely connected in the structural topology.

[0059] The present invention proposes a network hierarchical method, which is a condensed node extraction method based on semantic and structural features. It constructs three levels of different scales for the network and designs a hierarchical visualization method that supports large-scale complex semantic network exploration. In terms of technical means, a graph representation learning method is used to extract structural and semantic embedding vectors, and they are regarded as different view features of the network. On this basis, the present invention designs a multi-view clustering algorithm to extract condensed node clusters, and further aggregates them into mixed condensed nodes. Multiple condensed nodes are linked to each other to form a coarse-grained hierarchy.

[0060] In a hierarchical network, the essence of hierarchy definition is to find the mapping elements of actual nodes at different levels. Therefore, a basic aspect of describing a hierarchical network is to define and quantify the meaning and connection of cohesive nodes at different coarse-grained levels. Clarifying this point is also necessary for switching between levels in the system. The method proposed in the present invention completes the task of network hierarchization with complex semantic nodes. It not only considers the structural topology, but also analyzes the semantics of the text content contained in the node. In a large-scale semantic network, the hierarchical method not only relies on node shrinkage to reduce the visual complexity of the view layout, but also can maintain the continuation of the substructure content as much as possible while simplifying the structure. The present invention extracts a summary of the node that condenses its structural and semantic information. It filters out most of the information and retains some prominent features of the cohesive sub-nodes.

[0061] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the embodiments described in the specific implementation manners. Any other implementation manners derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A hierarchical method for large-scale complex semantic networks, characterized by: The steps include: S1. Network structure semantic information modeling: Extract node embedding vectors from the network to model the network structure and semantic information, preprocess the semantic information data of the network, and use the context-dependent network representation learning model CANE to generate a low-dimensional embedding representation with network node structure and semantic information to generate feature vectors for downstream tasks; S2. Multi-view clustering to extract mixed cohesive layers: Multi-view clustering is adopted to divide node clusters using semantic and structural information, and the node clusters corresponding to the hybrid cohesive nodes are further constructed with the help of the feature vectors obtained from the upstream task, namely the hybrid cohesive layer; S3, structural coacervate extraction: The community structure is used to further divide the network, so that the sub-cluster structure obtained by the new division tends to be dispersed, and a structural cohesive layer is obtained; S4. Hybrid Agglomerative Node Summary Extraction: Each mixed condensed node represents a semantic structure cluster. Three parts of information overview are generated around this cluster from the inside out. For the inside of the cluster, the summary shows the user the theme of the local network represented by the condensed node, and extracts the representative sub-nodes with important influence as the objects of special attention. For the outside of the cluster, the neighborhood level is displayed to reflect its closeness to other node themes in the global network. The embedding object of the network representation learning model CANE is to carry the nodes in the complex text network and generate the structural and semantic feature vectors with the same expression form. The node pairs linked by each edge are taken as input. The generated node semantic representation introduces the mutual attention mechanism according to the different node interaction neighbors, and has the characteristics of context-related. The multi-view clustering method reveals the intrinsic connection and complementary information between feature vectors obtained from upstream tasks from different perspectives, and seeks complementarity and consensus between views to assist clustering; In the network structure semantic information modeling step, the data preprocessing method is: The node text is case-standardized, grammatically eliminated, and stop words are removed. Stop words include functional words such as articles, prepositions, pronouns, and conjunctions.

2. The method for layering a large-scale complex semantic network according to claim 1, characterized in that: The multi-view clustering is implemented by the neighbor propagation algorithm, which is an unsupervised machine learning algorithm that does not require the specification of the number of clusters and creates clusters by sending and propagating messages between data points until convergence.

3. The hierarchical method of a large-scale complex semantic network according to claim 1, characterized in that: In the structural cohesive layer extraction step, the community structure is further divided into networks through a non-overlapping node community discovery algorithm.

Images