Graph representation learning device and method
The proposed graph representation learning method enhances accuracy and efficiency by partitioning graphs into subgraphs using community detection algorithms, addressing inefficiencies in existing GRL models and enabling scalable learning across diverse graph sizes and environments.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2026-04-02
AI Technical Summary
Existing graph representation learning (GRL) models face challenges in capturing detailed graph characteristics and changes, especially in large-scale graph processing, leading to inefficiencies and reduced accuracy due to high computational costs and inadequate partitioning strategies.
A graph representation learning device and method that employs a feature extraction unit, partitioning unit, and global graph configuration unit to select an accurate community detection algorithm, partition graphs into subgraphs, and learn representations at multiple levels, utilizing parallel processing for scalable and efficient learning.
Improves accuracy and learning efficiency of GRL models by considering graph characteristics, enabling scalable performance across varying graph sizes and deep learning environments.
Smart Images
Figure KR2024020187_02042026_PF_FP_ABST
Abstract
Description
Graph representation learning device and method
[0001] The present invention relates to a graph representation learning device and method, and in particular to a community detection-based multi-level graph representation learning device and method.
[0002] Graphs have been widely used to model and represent complex relationships in various domains, such as social networks, recommendation systems, and web graphs; however, as the complexity of graphs continues to increase, there is a demand for effective methods to extract meaningful information from these structures.
[0003] Graph Representation Learning (GRL) has emerged as a powerful solution for obtaining latent representations from graphs, and it can capture underlying structural information, node / edge attributes, and relationship dependencies within the graph. Furthermore, GRL can be utilized for various practical downstream tasks such as node classification, link prediction, and graph clustering.
[0004] Representative GRL approaches include random walk-based models and Graph Neural Network (GNN)-based models. Random walk-based models, such as DeepWalk and Node2Vec, use probabilistic models to generate sequences of nodes through random search within a graph. However, they cannot capture invisible nodes and cannot consider additional information such as node attributes or edge weights.
[0005] To overcome these limitations, GNN-based models such as Graph Convolutional Networks (GCNs) and GraphSAGE emerged, demonstrating better performance than random walk-based models. However, due to the inherent complexity of neural network structures, high computational costs are required, and the computational cost of model training increases exponentially as the size of the graph data grows.
[0006] Furthermore, existing GRL models proposed for large-scale graph processing focus primarily on efficient communication, lacking consideration for the accuracy of learned representations; additionally, existing community-based graph representation learning models apply partitioning only once regardless of the graph's size or characteristics, resulting in a problem where they do not consider detailed characteristics or changes within the graph.
[0007] Therefore, research is needed on methods to perform graph representation learning by considering the accuracy of graph representation as well as the detailed characteristics and changes of the graph.
[0008] The present invention was devised to solve the above-mentioned problems, and the objective of the present invention is to provide a graph representation learning device and method.
[0009] A graph representation learning device according to an embodiment of the present invention for achieving the above objective comprises: a feature extraction unit that extracts graph-based features from an input graph; a partitioning unit that selects a community detection algorithm with the highest accuracy of a graph representation learning (GRL) model using the extracted graph-based features and a pre-trained prediction model, and partitions the input graph into a plurality of subgraphs by applying the selected community detection algorithm; a global graph configuration unit that classifies the partitioned subgraphs into major communities and minor communities according to a predetermined node size, and integrates the subgraphs classified as major communities and the subgraphs classified as minor communities according to a predetermined method to form a major global graph and a minor global graph; and a graph representation learning unit that learns a major graph representation from the major global graph and learns a minor graph representation from the minor global graph through the GRL model.
[0010] A graph representation learning method according to an embodiment of the present invention for achieving the above objective comprises: a step of extracting graph-based features from an input graph; a step of selecting a community detection algorithm with the highest accuracy of a graph representation learning (GRL) model using the extracted graph-based features and a pre-trained prediction model, and partitioning the input graph into a plurality of subgraphs by applying the selected community detection algorithm; a step of classifying the partitioned subgraphs into major communities and minor communities according to a predetermined node size, and integrating the subgraphs classified as major communities and the subgraphs classified as minor communities according to a predetermined method to form a major global graph and a minor global graph; and a step of learning a major graph representation from the major global graph and learning a minor graph representation from the minor global graph through the GRL model.
[0011] According to one aspect of the present invention described above, by providing a graph representation learning device and method that perform partitioning by selecting an appropriate community detection algorithm according to graph characteristics, the accuracy and learning efficiency of existing GRL models can be simultaneously improved.
[0012] In addition, it is designed to take full advantage of parallel processing when learning graph representations, so it can be used scalably even as the size of graph data increases, and has the advantage of being universally applicable to various deep learning model structures and deployment environments.
[0013] FIG. 1 is a diagram illustrating the internal blocks of a graph representation learning device according to an embodiment of the present invention,
[0014] FIG. 2 is a diagram for explaining the overall operation of the graph representation learning device of FIG. 1,
[0015] FIG. 3a is a diagram showing the offline step of a graph representation learning method according to an embodiment of the present invention, and
[0016] FIG. 3b is a diagram showing the online step of a graph representation learning method according to an embodiment of the present invention, and
[0017] FIG. 4 is a diagram for explaining the operation of the partitioning unit of FIG. 1, and
[0018] FIG. 5 is a diagram for explaining the operation of the global graph generation unit of FIG. 1, and
[0019] FIG. 6 is a diagram for explaining the operation of the graph representation learning unit of FIG. 1, and
[0020] And, FIG. 7 is a flowchart illustrating a graph representation learning method of a graph representation learning device according to an embodiment of the present invention.
[0021] The following detailed description of the invention refers to the accompanying drawings, which illustrate specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that various embodiments of the invention are different but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in relation to one embodiment. It should also be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the following detailed description is not intended to be limiting, and the scope of the invention is limited only by the appended claims, including all equivalents to those claimed therein, provided appropriately described. Similar reference numerals in the drawings refer to the same or similar functions across various aspects.
[0022] The components according to the present invention are defined by functional distinction rather than physical distinction, and can be defined by the functions each performs. Each component may be implemented as hardware or as program code and processing units that perform each function, and the functions of two or more components may be included and implemented in a single component. Therefore, it should be noted that the names assigned to the components in the following embodiments are not intended to physically distinguish each component but are assigned to imply the representative function performed by each component, and that the technical concept of the present invention is not limited by the names of the components.
[0023] This invention proposes a methodology for learning multi-level graph representations based on community detection to improve both the accuracy and learning efficiency of existing GRL models. To this end, a community detection algorithm best suited to the graph is selected through a prediction model, and the graph is recursively partitioned into multiple subgraphs. Then, using a global graph that reflects the topology of the entire graph, minor graph representation learning and major graph representation learning classified according to the size of the subgraphs are performed to learn the representation of the entire graph.
[0024] Preferred embodiments of the present invention will be described in more detail below with reference to the drawings.
[0025] FIG. 1 is a diagram illustrating the internal blocks of a graph representation learning device according to an embodiment of the present invention, FIG. 2 is a diagram explaining the overall operation of the graph representation learning device of FIG. 1, FIG. 3a is a diagram showing the offline step of a graph representation learning method according to an embodiment of the present invention, FIG. 3b is a diagram showing the online step of a graph representation learning method according to an embodiment of the present invention, FIG. 4 is a diagram explaining the operation of the partitioning unit of FIG. 1, FIG. 5 is a diagram explaining the operation of the global graph generation unit of FIG. 1, and FIG. 6 is a diagram explaining the operation of the graph representation learning unit of FIG. 1.
[0026] The illustrated graph representation learning device includes a feature extraction unit (110), a partitioning unit (120), a global graph configuration unit (130), and a graph representation learning unit (140).
[0027] The feature extraction unit (110) extracts graph-based features from the input graph.
[0028] The partitioning unit (120) uses the graph-based features extracted by the feature extraction unit (110) and a pre-trained prediction model to select the community detection algorithm with the highest accuracy of the Graph Representation Learning (GRL) model, and applies the selected community detection algorithm to partition the input graph into multiple subgraphs. Figure 2 shows an example of partitioning the input graph into two stages, and it can be seen that the final subgraph that can no longer be partitioned exists at a different level, as indicated by the dotted line.
[0029] The global graph configuration unit (130) classifies the subgraphs partitioned through the partitioning unit (120) into a major community and a minor community according to a predetermined node size, and unifies the subgraphs classified into the major community and the subgraphs classified into the minor community according to a predetermined method to form a major global graph and a minor global graph. Here, the node size is defined as the number of nodes, and in FIG. 2, subgraphs with fewer than 3 nodes are assumed to be the minor community, and the remaining subgraphs are assumed to be the major community.
[0030] The graph representation learning unit (140) learns a major graph representation from the major global graph and learns a minor graph representation from the minor global graph through the GRL model.
[0031] Meanwhile, the graph representation learning method of the graph representation learning device according to an embodiment of the present invention can be divided into an offline phase and an online phase. Here, the offline phase is a process of pre-training a prediction model using a sample dataset, and the online phase is a process of performing graph representation learning after performing multi-level partitioning using the pre-trained prediction model.
[0032] First, looking at the offline stage through Fig. 3a, the graph representation learning device uses a pre-prepared training dataset G train By applying the neighbor sampling method, which samples a fixed-size set of neighbors, to multiple sample data G' trainIt generates a fixed-size node from a specific graph, and then generates each sampled graph by adding 1-hop and 2-hop neighbors with the same number of neighbors to the sampled node. The reason for applying the neighbor sampling method is to ensure that each sample data retains the structural information of the original graph.
[0033] Then, the graph representation learning device extracts graph-based features from the generated sample data, which can be extracted without performing graph partitioning. The graph-based features are defined as statistical features and structural features, the statistical features include the number of nodes and edges, the dimension of node attributes, and statistical values of node degrees (mean, maximum, minimum, standard deviation), and the structural features include density, degree-based centrality (mean, maximum, minimum, standard deviation), and clustering coefficients.
[0034] Subsequently, the graph representation learning device generates a prediction model for each community detection algorithm, and pre-trains the prediction model by setting the accuracy of the GRL model measured after performing partitioning by applying at least one community detection algorithm to the sample data using the extracted graph-based features as input variables as output variables. At this time, the accuracy of the GRL model measured when partitioning is not performed is also pre-trained, and the prediction model may be generated according to the type of GRL model and the downstream task for measuring the accuracy of the GRL model. In the present invention, not performing partitioning is defined as a "No Partitioning" algorithm.
[0035] Next, looking at the online step through FIG. 3b, the feature extraction unit (110) input graph G inputGraph-based features are extracted from, and the partitioning unit (120) uses the extracted graph-based features and a prediction model P that has been pre-trained through an offline step. S Select the community detection algorithm with the highest accuracy of the GRL model using [this].
[0036] Then, the partitioning unit (120) applies the selected community detection algorithm to the input graph G input The partitioning unit (120) partitions the data into multiple subgraphs. The partitioning unit (120) then performs recursive partitioning for each subgraph created thereafter, and this partitioning operation is terminated when a no-partitioning algorithm is selected or when the number of nodes in the corresponding subgraph is less than a predetermined number. In FIG. 4, the predetermined number ρ is set to 3, so it can be seen that partitioning is terminated for subgraphs with fewer than 3 nodes and subgraphs for which a no-partitioning algorithm is selected, and partitioning is performed only on the remaining subgraphs.
[0037] The global graph configuration unit (130) classifies subgraphs partitioned into multiple levels into a major community and a minor community according to a predetermined node size, and integrates the subgraphs classified into the major community and the subgraphs classified into the minor community according to a predetermined method to form a major global graph and a minor global graph. Here, the major community refers to a significant subgraph of a certain size or larger, and the minor community refers to a small subgraph of a size less than a certain size.
[0038] To explain the above global graph configuration step in more detail with reference to FIG. 5, the global graph configuration part (130) is the major community C majorCreate a major community graph by abstracting subgraphs of different levels classified as to a specific level, for example, the top level, and minor community C minor Subgraphs of different levels classified as such are level-free merged to create a minor community graph, and the major community graph and the minor community graph are concatenated to obtain a unified community graph.
[0039] Then, the global graph configuration unit (130) abstracts a subgraph classified as a major community among the integrated community graphs into at least one super node, and connects the at least one super node with a subgraph classified as a minor community to form a minor global graph.
[0040] In addition, the community with the most nodes among the 1-hop neighbors of the subgraph classified as the minor community is merged with the subgraph classified as the minor community to create at least one representative node, and the at least one representative node is connected to form a Major Global Graph.
[0041] To explain the process of generating the representative node through the merge process shown in Fig. 5, the subgraph classified as a minor community, indicated by Node 1, is merged with the community with the most nodes among its 1-hop neighbors, i.e., the community indicated by Node 3, to generate Node 4, and the generated Node 4 is merged with the subgraph classified as a minor community, indicated by Node 2, to generate Representative Node 6. Here, the number indicated for each node represents the number of nodes included in it, and through this, not only super nodes but also the same minor communities can be merged. For example, if the only neighbors of a minor community are other minor communities, the two can be merged together.
[0042] In this case, for both the major global graph and the minor global graph, the attribute of each node is assigned a value aggregated as the average or mode of all nodes belonging to that node.
[0043] Meanwhile, the present invention proposes minor graph representation learning for the learning of minor communities and major graph representation learning for the learning of major communities. Since minor communities do not possess sufficient size, the accuracy of graph representation may decrease if graph representation learning is performed independently. Therefore, minor graph representations are learned by integrating them with other minor communities into a single graph and using a minor global graph. On the other hand, since major communities are large enough to contain meaningful information, major graph representations are learned by applying a major global graph to each major community.
[0044] As explained in more detail through FIG. 6, the graph representation learning unit (140) connects the major global graph to the subgraph classified into each major community, and then performs major graph representation learning in parallel to produce major embeddings of nodes belonging to the major community.
[0045] In addition, the graph representation learning unit (140) performs minor graph representation learning from the minor global graph to produce minor embeddings of nodes belonging to the minor community.
[0046] Finally, the graph representation learning unit (140) connects the major embedding and the minor embedding to generate a final embedding.
[0047] FIG. 7 is a flowchart illustrating a graph representation learning method of a graph representation learning device according to an embodiment of the present invention.
[0048] A graph representation learning device extracts graph-based features from an input graph. (S701)
[0049] Then, the graph representation learning device selects the community detection algorithm with the highest accuracy of the GRL model using the graph-based features extracted in S701 and a pre-trained prediction model, and partitions the input graph into multiple sub-graphs by applying the selected community detection algorithm. (S703)
[0050] Then, the graph representation learning device classifies the subgraphs partitioned in S703 into major communities and minor communities according to a predetermined node size, and integrates the subgraphs classified into major communities and the subgraphs classified into minor communities according to a predetermined method to form a major global graph and a minor global graph. (S705)
[0051] Subsequently, the graph representation learning device learns a major graph representation from the major global graph and learns a minor graph representation from the minor global graph through a GRL model. (S707)
[0052] The graph representation learning method of the present invention, as described above, can be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., either individually or in combination.
[0053] The program instructions recorded on the above-mentioned computer-readable recording medium may be those specifically designed and configured for the present invention, or they may be those known and available to those skilled in the art of computer software.
[0054] Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, and flash memory.
[0055] Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.
[0056] Although various embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention.
[0057] (Explanation of symbols)
[0058] 110: Feature extraction unit
[0059] 120: Partitioning section
[0060] 130: Global Graph Components
[0061] 140: Graph Representation Learning Section
Claims
1. A feature extraction unit that extracts graph-based features from an input graph; A partitioning unit that selects a community detection algorithm with the highest accuracy of a Graph Representation Learning (GRL) model using the extracted graph-based features and a pre-trained prediction model, and partitions the input graph into multiple subgraphs by applying the selected community detection algorithm; A global graph configuration unit that classifies the partitioned subgraphs into a major community and a minor community according to a predetermined node size, and unifies the subgraphs classified into the major community and the subgraphs classified into the minor community according to a predetermined method to form a major global graph and a minor global graph; and A graph representation learning device comprising: a graph representation learning unit that learns a major graph representation from the major global graph and learns a minor graph representation from the minor global graph through the above GRL model.
2. In Paragraph 1, The above prediction model is, A graph representation learning device that is pre-trained by setting graph-based features extracted from sample data sampled from a pre-prepared training dataset as input variables, applying at least one community detection algorithm to the sample data to perform partitioning, and then setting the accuracy of the measured GRL model as output variables.
3. In Paragraph 1, The above global graph component is, The subgraph classified as the major community is abstracted into at least one super node, and the minor global graph is constructed by connecting the at least one super node and the subgraph classified as the minor community. A graph representation learning device that generates at least one representative node by merging the community having the most nodes among the 1-hop neighbors of the subgraph classified as the minor community with the subgraph classified as the minor community, and constructs the major global graph by connecting the at least one representative node.
4. In Paragraph 1, The above graph representation learning unit is, After connecting the above major global graph to the subgraphs classified into each major community, major graph representation learning is performed in parallel to calculate the major embeddings of the nodes belonging to the above major communities, and Minor graph representation learning is performed from the above minor global graph to calculate minor embeddings of nodes belonging to the above minor community, and A graph representation learning device that generates a final embedding by concatenating the above major embeddings and minor embeddings.
5. Step of extracting graph-based features from the input graph; A step of selecting a community detection algorithm with the highest accuracy of a Graph Representation Learning (GRL) model using the extracted graph-based features and a pre-trained prediction model, and partitioning the input graph into multiple subgraphs by applying the selected community detection algorithm; A step of classifying the partitioned subgraphs into a major community and a minor community according to a predetermined node size, and unifying the subgraphs classified into the major community and the subgraphs classified into the minor community according to a predetermined method to form a major global graph and a minor global graph; and A graph representation learning method comprising the step of learning a major graph representation from the major global graph and learning a minor graph representation from the minor global graph through the above GRL model.