KR20210089033A - Hierarchical clustering method and system for directivity graph data - Google Patents

Abstract

A hierarchical clustering method of directed graph data in accordance with the present invention comprises the steps of: obtaining directed graph data; creating a node relevance information cluster from node and link information constituting each graph; partitioning the node relevance information cluster by a max-flow min-cut theorem; and checking whether the divided structure of the node relevance information cluster reaches a desired tree structure. Here, in the step of creating the node relevance information cluster, in the case of a graph in which a weighting factor is given to a link or an edge connecting two nodes, the weighting factor is used as it is, and if no weighting factor is given to all links/edges, a weighting factor of the basic unit amount can be given to all links/edges. The present invention can cluster the directed graph data with a tree structure having high application versatility.

Description

HIERARCHICAL CLUSTERING METHOD AND SYSTEM FOR DIRECTIVITY GRAPH DATA

The present invention relates to a hierarchical clustering method and system for clustering directed graph data into a tree structure.

Information Extraction is a subfield of computer science that aims to make machines acquire knowledge automatically by extracting structured information from unstructured data. Until the early 2000s, information extraction technology focused on extracting answers to specific queries or a large number of examples corresponding to predefined relationships using hand-crafted extraction rules or learning examples. For example, research such as extracting a large number of tuples of the form (author or author) from a web document through a pattern, or automatically fetching a price from each site for a specific book.

Clustering is a kind of data processing technology that classifies given data into several related clusters, and is a kind of preprocessing process widely used in numerous data analysis including machine learning. Representative clustering techniques include k-means clustering, hierarchical clustering, and spectral clustering.

Most clustering problems focus on clustering single data, but clustering multiple related data at once is also an important issue.

As one of the effective data expression methods to reflect the meaning of data, the knowledge graph is being used and attracting attention in many applications such as knowledge extraction, Q&A, and information retrieval.

In general, the knowledge graph is composed of a set of triples <h,r,t> called facts (facts), and the head entity (h) and the tail entity (t) are connected by a semantic relationship (r). Various knowledge graphs such as DBpedia, WikiData, Freebase, WordNet, and Google Knowledge Graph have been developed for massive heterogeneous knowledge integration and interoperability.

Proposed techniques for clustering nodes constituting a graph by analyzing graph data are as follows.

1 is an automatic clustering technique of word meaning using a word association graph, which is applied to word co-occurrence information extracted from a web document to generate a cluster of semantically similar related words through a graph-based clustering technique. They are created as a group of meanings. In addition, a relative cluster evaluation index is defined and the fitness of the generated cluster is evaluated by applying it.

2 is a configuration for generating frequent subgraphs for graph classification. Instead of generating all frequent subgraphs of graphs constituting a graph database, similar frequent subgraphs are grouped into similar groups based on structural similarity, and in each similarity group We propose a configuration that generates a frequent subgraph with the highest classification power.

3 is a flowchart of a multi-step clustering method of documents using an ontology. By clustering documents using an ontology including the main words of the document, similar documents according to the subject words of the document are effectively grouped together.

4 is an information extraction augmentation system through implicit relationship inference between entities. From a set of tuples extracted from OIE, a new information tuple that is not explicitly described in an existing document is generated through implicit relationship inference between entities to enhance the information extraction result. We present a system that makes

Figure 5 is a community detection method, in a multilayer graph, deriving the importance of each layer for each layer, flattening the multilayer graph by applying the derived importance for each layer, and detecting a community using the flattened multilayer graph represent the steps.

6 is a tree-based gradual clustering apparatus using a dependent normalized random measure. In order to efficiently cluster several related data without prior information on the number of clusters, a tree-based multiple We present a device for gradual clustering rather than hierarchically related data.

FIG. 7 is a community detection framework device, and presents a framework capable of generating an object relationship diagram showing a community structure well by using spatial transformation and detecting a community based on the object relationship diagram or object similarity information.

However, all of the above-mentioned conventional techniques have the following problems.

A multi-level clustering method of documents using ontology has been proposed, but it is not a method of clustering documents into a tree structure having a hierarchical structure. Therefore, this method proposed in the prior art does not have the advantage of a tree type data structure.

A method of detecting a community by deriving the importance of each layer from a given multi-layer graph has been proposed, but since this is community detection in a given situation in a multi-layer graph already structured hierarchically, only limited application fields exist.

In addition, a tree-based gradual clustering method has been proposed, but this method is an algorithm that gradually constructs data elements in a tree form, and each element forms a node of a tree structure. Therefore, even when there is no hierarchical order among elements constituting data, there is a disadvantage in constructing a tree by giving artificial hierarchical order.

In conclusion, when there is directed graph data, a method and algorithm for generating tree-structured data by hierarchically clustering nodes constituting the graph, and a computer implementing the same The system has not been presented.

Republic of Korea Publication No. 10-2018-0093158

An object of the present invention is to provide a hierarchical clustering method and/or system for directed graph data capable of clustering directed graph data into a tree structure with high application versatility.

More specifically, in the present invention, when an arbitrary directed graph is given, by using information of nodes and links (link/edge) constituting the graph, nodes with high relevance are clustered, An object of the present invention is to provide a hierarchical clustering method and/or system of directed graph data that hierarchically forms a cluster to which each node belongs by hierarchically generating such clusters.

The present invention can be applied well even when there is no hierarchical order between nodes constituting the graph, and is a hierarchical clustering of directed graph data that can automatically generate a hierarchical cluster tree from a large amount of graph data using a computer system. It is intended to provide a method and/or system.

A hierarchical clustering method of directed graph data according to an aspect of the present invention includes: obtaining directed graph data; creating a node relevance information cluster from node and link information constituting each graph; partitioning the node relevance information cluster by maximum flow minimum truncation; and checking whether the divided structure of the node relevance information cluster reaches a desired tree structure.

Here, in the step of creating the node relevance information cluster, in the case of a graph in which weights are given to links or edges connecting two nodes, the weights are used as they are, and when no weights are given to all links/edges, all Links/edges can be given a weight of the basic unit amount.

Here, the step of dividing the node relevance information cluster may include: selecting a source node and a sink node from the created node relevance information cluster; determining guidelines of maximum flow and minimum truncation applied in dividing the node relevance information cluster based on the source node and the sink node; and dividing the node relevance information cluster according to the determined guideline.

Here, in the step of selecting the source node and the sink node, the node having the largest sum of weights of edges leaving the node and the node having the largest sum of weights of edges entering the node may be designated as the source node and the sink node, respectively.

Here, in the step of dividing the node relevance information cluster, the maximum flow condition or the minimum cutting condition is applied to perform division only when the expected maximum flow degree or minimum cutting degree reaches or exceeds each predetermined criterion, In the step of checking whether the desired tree structure is reached, it can be confirmed that the partition is no longer possible.

A hierarchical clustering system for directed graph data according to another aspect of the present invention includes: a directed graph data obtaining unit for obtaining directed graph data; a node relevance information cluster creation unit for creating a node relevance information cluster from the node and link information constituting each graph; a source/sink node selection unit for selecting a source node and a sink node from the created node relevance information cluster; a guideline determining unit for determining a maximum flow and minimum truncation guidelines applied in dividing the node relevance information cluster based on the source node and the sink node; and a partitioning and tree forming unit for partitioning the node relevance information cluster according to the determined guideline.

Here, the node relevance information cluster creation unit uses the weights as it is in the case of a graph in which weights are given to links or edges connecting two nodes, and when no weights are assigned to all links/edges, all links/edges can be given a weight of the basic unit amount.

Here, the source/sink node selector may designate a node having the largest sum of weights of edges leaving the node and a node having the largest sum of weights of edges entering the node as a source node and a sink node, respectively.

Here, the division and tree forming unit performs division only when the maximum flow or minimum cutting degree expected at the time of division by applying the maximum flow condition or the minimum cutting condition reaches or exceeds each predetermined criterion, and further division is possible By confirming that it does not, the tree structure can be confirmed.

If the hierarchical clustering method and/or system of the directed graph data of the present invention having the above configuration is implemented, there is an advantage in that the directed graph data can be clustered into a tree structure with high application versatility.

The hierarchical clustering method and/or system of directed graph data of the present invention can be well applied even when there is no hierarchical order between nodes constituting the graph, and can be automatically hierarchical from large amounts of graph data using a computer system. It has the advantage of being able to create a cluster tree.

The hierarchical clustering method and/or system of directed graph data of the present invention can automatically form a hierarchical cluster of nodes by utilizing the correlation of nodes constituting the graph when there is arbitrary directed graph data, Since such a hierarchical cluster has a tree structure, there is an advantage in quickly identifying and utilizing the clusters to which each node belongs.

For example, when there is document data constituting Wikipedia, each Wikipedia document is viewed as a node, and if hyperlinks between Wikipedia documents are placed as an edge, a large amount of concepts (Wikipedia documents/pages) are automatically combined with similarity. You can create a hierarchical classification system by clustering according to /relevance.

1 is a conceptual diagram illustrating an automatic clustering technique of word meanings using a word association graph.
2 is a block diagram showing a configuration for generating a frequent partial graph for graph classification.
3 is a flowchart of a multi-step clustering method of documents using an ontology.
4 is a block diagram illustrating an information extraction augmentation system through implicit relationship inference between entities.
5 is a conceptual diagram illustrating a method of detecting a community using a flattened multi-layer graph.
6 is a block diagram illustrating a tree-based gradual clustering device using a dependent normalized random measure.
7 is a conceptual diagram illustrating a framework capable of generating an object relationship diagram showing a community structure well by using spatial transformation and detecting a community based on object relationship diagram or object similarity information.
8 is a flowchart illustrating a hierarchical clustering method of directed graph data according to an embodiment of the present invention.
Fig. 9 is a data structure diagram visually illustrating an example of any given graph;
10 is a flowchart illustrating a detailed configuration of a step of dividing the node relevance information cluster of FIG. 8;
11 is a data configuration diagram illustrating a state in which the node configuration of FIG. 9 is divided by the hierarchical clustering method of FIG. 8;
12 is a conceptual diagram illustrating a hierarchical tree formation process according to hierarchical clustering of directed graph data according to the spirit of the present invention.
13 is an algorithm expression in which the hierarchical clustering method of directed graph data according to the spirit of the present invention described above is implemented with pseudocode.
14 is a block diagram illustrating a hierarchical clustering system of directed graph data according to an embodiment of the present invention.

In describing the present invention, terms such as first, second, etc. may be used to describe various components, but the components may not be limited by the terms. The terms are only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When a component is referred to as being connected or connected to another component, it may be directly connected or connected to the other component, but it can be understood that other components may exist in between. .

The terms used herein are used only to describe specific embodiments, and are not intended to limit the present invention. The singular expression may include the plural expression unless the context clearly dictates otherwise.

In this specification, the terms include or include are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, and includes one or more other features or numbers, It may be understood that the existence or addition of steps, operations, components, parts or combinations thereof is not precluded in advance.

In addition, shapes and sizes of elements in the drawings may be exaggerated for clearer description.

When an arbitrary directed graph is given, the present invention clusters highly related nodes using information on nodes and links (link/edge) constituting the graph, so that each node belongs We propose a new method for hierarchically forming clusters.

8 is a flowchart illustrating a hierarchical clustering method of directed graph data according to an embodiment of the present invention.

The illustrated hierarchical clustering method of directed graph data includes: obtaining directed graph data (S100); creating a node relevance information cluster from node and link information constituting each graph (S200); dividing the node relevance information cluster by maximum flow minimum truncation (S300); and confirming whether the divided structure of the node relevance information cluster reaches a desired tree structure ( S400 ).

The directed graph data may include a large number of directed graphs, and each graph may have a triple data set similar to a general graph. Meanwhile, for directionality, each graph has node information and link or edge information connecting nodes, but information indicating direction and/or weight may be assigned to the link or edge.

Next, individual graphs and the process of creating a cluster of node relevance information for them ( S200 ) will be described with specific examples. A visual representation of an example of any given graph is shown in FIG. 9 . In FIG. 9, edge weights are omitted.

The relevance of the nodes in the graph is determined by obtaining information from the links or edges connecting the nodes. In the case of a graph in which weights are given to links or edges connecting two nodes, the weights are used as they are, and if no weights are assigned to all links/edges, the basic unit amount (e.g., , 1) is weighted. The higher the weight of the link/edge connecting two nodes, the higher the relevance of the two nodes is considered.

The illustrated node relevance information cluster may be stored in a DB or a storage device in a form according to a known clustering information storage algorithm.

Next, the step of dividing the node relevance information cluster ( S300 ) will be described with a specific example.

The well-known graph-cut theory, max-flow min-cut theorem, is iteratively applied to form clusters in a hierarchical tree structure from large clusters at high levels to small clusters at low levels. Existing nodes constituting the graph constitute the last leaf node (leaf node) of the tree structure of newly formed clusters.

When dividing the nodes of the graph into different clusters by changing to the maximum flow minimum truncation problem, the weight of the graph link/edge is treated as the capacity of the link/edge. Algorithms to solve the max-flow min-cut theorem can use well-known linear programming optimization methods.

FIG. 10 shows a detailed configuration of the step of dividing the node relevance information cluster of FIG. 8 , and FIG. 11 shows a state in which the node configuration of FIG. 9 is divided by the hierarchical clustering method of FIG. 8 .

Next, detailed processes of the step (S300) of dividing the node relevance information cluster according to the flow shown in FIG. 10 will be described.

selecting a source node and a sink node from the node relevance information cluster created in step S200 (S310); determining guidelines of maximum flow and minimum truncation applied in dividing the node relevance information cluster based on the source node and the sink node (S320); and dividing the node relevance information cluster according to the determined guideline (S330).

In order to solve one maximum flow minimum truncation problem, a source node and a sink node of a given graph should be determined (S310). There may be several methods for selecting a source node and a sink node in an arbitrary given graph, but in the present invention, the node having the largest sum of weights of edges leaving the node and the node having the largest sum of weights of edges entering the node are selected, respectively. We propose a method of designating a source node and a sink node.

Set the source node and sink node in the method suggested above, and find the maximum flow (or minimum truncation) as shown in the figure below. As shown in the figure below, by solving one maximum flow problem, the nodes of the graph can be divided into two different clusters.

On the other hand, the maximum flow minimum cutting algorithm has two conditions, the maximum flow condition and the minimum cutting condition, even in the name. Various cases may exist in the guidelines for applying the two conditions when cutting (segmenting) the graph cluster in step S300. In the illustrated step S320, one of the various cases is selected, and in the illustrated step S330, the determined guidelines are used. Partition graph clusters by applying maximum flow conditions and/or minimum cut conditions.

For example, in the case of a field in which maximum flow maintenance is important, in the step S300, only the maximum flow condition is applied to divide, or the maximum flow condition is preferentially divided to apply the division, but when the maximum flow rate expected at the time of division does not meet the standard, the minimum It can be divided by applying cutting conditions.

For example, in the case of a field in which it is important to ensure minimum cutting, in the step S300, only the minimum cutting condition is applied to divide, or preferentially apply the minimum cutting condition to divide, but when the expected minimum cutting degree does not meet the standard, the maximum It can be segmented by applying flow conditions.

For example, in the step S300, the two conditions may be alternately applied, such as applying the maximum flow condition in the first division and applying the minimum flow condition in the next division.

For example, in the step S300, the maximum flow condition or the minimum cutting condition may be applied to perform division only when the expected maximum flow level or minimum cutting level reaches or exceeds each predetermined criterion.

Next, the step of checking whether the desired tree structure is reached ( S400 ) will be described with a specific example.

As a result of performing the above-described step S300, the nodes of the graph are divided into two different clusters, and each cluster can be treated as an individual graph, and the cluster can be repeatedly divided into smaller unit clusters. By repeating this process, clusters having a tree structure of an appropriate size can be formed. The concept of such a hierarchical tree formation process is illustrated in FIG. 12 .

Criteria for checking whether a desired tree structure is reached in step S400 may vary according to implementations. For example, in the case of FIG. 12 , the tree is formed under the condition that only three or less nodes belong to the finally divided tree branch cluster.

For example, in step S400, it may be checked whether a condition for limiting the number of layers of the tree to a predetermined number is reached.

For example, in the step S300, if the maximum flow condition or the minimum cutting condition is applied and the division is performed only when the expected maximum flow or minimum cutting degree reaches or exceeds each predetermined criterion, in the S400 step, the It may be performed by confirming that the division is not possible.

In the above-described algorithm, a tree-structured cluster (or category) can be created step by step from a high-level cluster to a low-level cluster by writing a function in a recursion method.

For example, in the step S400, the stopping criterion of the recursive call of the above-described algorithm is set according to the application field, thereby preventing the case of infinitely repeatedly calling the recursive function. As the stopping condition, the number of nodes in the given graph is less than a preset reference point, or other similar criteria may be set.

13 shows an algorithm in which the hierarchical clustering method of directed graph data according to the spirit of the present invention described above is implemented with pseudocode. Here, V denotes a set of nodes in the graph, and E denotes a set of edges/links.

14 is a block diagram illustrating a hierarchical clustering system of directed graph data according to an embodiment of the present invention.

The illustrated hierarchical clustering system 100 of directed graph data includes: a directed graph data obtaining unit 110 for obtaining directed graph data (information); a node relevance information cluster creation unit 120 for creating a node relevance information cluster from the node and link information constituting each graph; a source/sink node selection unit 131 for selecting a source node and a sink node from the created node relevance information cluster; a guideline determining unit 132 for determining a guideline for maximum flow and minimum truncation applied in dividing the node relevance information cluster based on the source node and the sink node; and a partitioning and tree forming unit 133 for partitioning the node relevance information cluster according to the determined guideline.

The directed graph data may include a large number of directed graphs, and each graph may have a triple data set similar to a general graph. Meanwhile, for directionality, each graph has node information and link or edge information connecting nodes, but information indicating direction and/or weight may be assigned to the link or edge.

The direction graph data acquisition unit 110 may perform step S100 of FIG. 8 described above in such a way that the direction graph data is transmitted from an external directed graph information collection DB or the like using wired/wireless data communication. Depending on the implementation, the directed graph data obtaining unit 110 may store the received directed graph data in the storage unit 160 formed in the hierarchical clustering system 100 of directed graph data.

The node relevance information cluster creation unit 120 may perform the above-described step S200 of FIG. 8 to create a node relevance information cluster configured as shown in FIG. 9 and store it in the illustrated storage unit 160 .

The source/sink node selection unit 131 reads out the node relevance information cluster stored in the illustrated storage unit 160, and determines a source node and a sink node, which are the basis for a subsequent division operation, as shown in FIG. Step S310 may be performed.

The guide determination unit 132 determines guidelines for maximum flow conditions and minimum cutting conditions in various cases that can be applied in performing the above-described step S320, determined according to the type of circular graph data, or by a user in advance. It can be performed by reading the specified and recorded instructions.

The division and tree forming unit 133 performs steps S330 and S400 described above. The division and tree forming unit 133 temporarily stores in the storage unit 160 showing the cluster cluster divided as a result of the execution of step S330, and stores the tree structure determined in step S400 and the cluster cluster divided accordingly. It may be finally stored in the illustrated storage unit 160 .

For example, the division and tree forming unit 133 performs division only when the maximum flow or minimum cutting degree expected at the time of division by applying the maximum flow condition or the minimum cutting condition reaches or exceeds each predetermined criterion, By confirming that it is no longer splittable, the tree structure can be finally confirmed.

Those skilled in the art to which the present invention pertains should understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof, so the embodiments described above are illustrative in all respects and not restrictive. only do The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

For example, a clustering algorithm according to the spirit of the present invention may be applied to each hardware configuration presented in the prior art shown in FIGS. 1 to 7 , which also falls within the scope of the present invention.

100: hierarchical clustering system of directed graph data
110: direction graph data acquisition unit
120: node relevance information cluster creation unit
131: source/sink node selection unit
132: guide decision unit
133: segmentation and tree forming part
160: storage

Claims (9)

obtaining directed graph data;
creating a node relevance information cluster from node and link information constituting each graph;
partitioning the node relevance information cluster by maximum flow minimum truncation; and
Checking whether the divided structure of the node relevance information cluster reaches a desired tree structure
A hierarchical clustering method of directed graph data comprising
According to claim 1,
In the step of creating the node relevance information cluster,
In the case of a graph in which weights are given to links or edges connecting two nodes, the weights are used as they are, and when no weights are assigned to all links/edges, the weight of the basic unit amount is given to all links/edges. A hierarchical clustering method of graph data.
According to claim 1,
The step of dividing the node relevance information cluster comprises:
selecting a source node and a sink node from the created node relevance information cluster;
determining guidelines of maximum flow and minimum truncation applied in dividing the node relevance information cluster based on the source node and the sink node; and
Segmenting the node relevance information cluster according to the determined guideline
A hierarchical clustering method of directed graph data comprising
4. The method of claim 3,
In the step of selecting the source node and the sink node, the node having the largest sum of weights of the edges leaving the node and the node having the largest sum of the weights of the edges entering the node are designated as the source node and the sink node, respectively. hierarchical clustering method.
4. The method of claim 3,
In the step of dividing the node relevance information cluster, the division is performed only when the maximum flow degree or the minimum cutting degree expected at the time of division by applying the maximum flow condition or the minimum cutting condition reaches or exceeds each predetermined criterion,
In the step of checking whether the desired tree structure is reached, a hierarchical clustering method of directed graph data to confirm that it is no longer divisible.
a directed graph data obtaining unit for obtaining directed graph data;
a node relevance information cluster creation unit for creating a node relevance information cluster from the node and link information constituting each graph;
a source/sink node selection unit for selecting a source node and a sink node from the created node relevance information cluster;
a guideline determining unit for determining a maximum flow and minimum truncation guidelines applied in dividing the node relevance information cluster based on the source node and the sink node; and
A partitioning and tree forming unit that divides the node relevance information cluster according to the determined guideline.
A hierarchical clustering system of directed graph data comprising a.
7. The method of claim 6,
The node relevance information cluster creation unit,
In the case of a graph in which weights are given to links or edges connecting two nodes, the weights are used as they are, and when no weights are assigned to all links/edges, the weight of the basic unit amount is given to all links/edges. A hierarchical clustering system for graph data.
7. The method of claim 6,
The source/sink node selection unit,
A hierarchical clustering system for directed graph data that designates the node with the largest sum of weights of outgoing edges and the node with the largest sum of weights of incoming edges as source and sink nodes, respectively.
7. The method of claim 6,
The division and tree forming unit,
By applying the maximum flow condition or the minimum cutting condition, the division is performed only when the expected maximum flow degree or the minimum cutting degree reaches or exceeds each predetermined criterion at the time of division,
A hierarchical clustering system of directed graph data that establishes a tree structure upon confirmation that it is no longer divisible.

Images