KR-20260063660-A - GRAPH CLASSIFICATION METHOD AND APPARATUS BASED ON AUGMENTATION-AWARE REPRESENTATION LEARNING

Abstract

A graph classification method and a graph classification device are presented. The graph classification method performed by the graph classification device includes the step of classifying one or more graphs by inputting one or more graphs into a graph classification model comprising an encoder that outputs a representation of a graph and a classifier connected to the encoder that predicts the class of the graph based on the representation of the graph. The graph classification model is trained to recognize the augmentation difference between the original graph and the augmented graph using a graph representation learning model in which the representations of the original graph and the augmented graph among the one or more graphs are input.

Inventors

• 강유
• 김민준
• 최재현
• 이승주
• 정진홍

Assignees

• 서울대학교산학협력단

Dates

Publication Date

20260507

Application Date

20241030

Claims (10)

1. In a graph classification method performed by a graph classification device, The method includes the step of classifying one or more graphs by inputting one or more graphs into a graph classification model comprising an encoder that outputs a representation of a graph and a classifier connected to the encoder that predicts the class of the graph based on the representation of the graph. The above graph classification model is, A graph classification method that learns to approximate the distance between the original graph and the augmented graph by taking the representation of the original graph and the representation of the augmented graph as inputs, and learns to recognize the augmentation difference between the original graph and the augmented graph using the output of a graph representation learning model and the distance.
2. In paragraph 1, The above graph representation learning model is, A graph classification method that receives data combining the representation of the original graph and the representation of the augmented graph, and approximates the distance between the original graph and the augmented graph by the representation of the original graph and the representation of the augmented graph.
3. In paragraph 1, The above graph classification model is, A graph classification method that learns to recognize the augmentation difference by aligning the difference in graph levels according to the distance between the original graph and the augmented graph in a graph space where the original graph and the augmented graph are mapped, with the representation of the original graph and the representation of the augmented graph in a representation space where the representation of the original graph and the representation of the augmented graph are mapped, through the graph representation learning model.
4. In paragraph 1, The above graph classification model is, A graph classification method that, through the graph representation learning model, injects the characteristic distance between the features of the original graph and the features of the augmented graph into the relationship between the representation of the original graph and the representation of the augmented graph in a feature space where the features of the vertices of the original graph and the features of the vertices of the augmented graph are mapped.
5. In paragraph 1, The above graph classification model is, A graph classification method that, through the above graph representation learning model, injects the structural distance between the structure of the original graph and the structure of the augmented graph in the structure space where the original graph is mapped, into the relationship between the representation of the original graph and the representation of the augmented graph.
6. In paragraph 1, The above graph classification model is, A graph classification method in which the class label based on the probability predicted for the original graph and the class label based on the probability predicted for the augmented graph are learned to be the same.
7. In paragraph 1, The above graph classification model is, A graph classification method learned according to (i) a first loss function defined to classify the original graph and the augmented graph, (ii) a second loss function defined to recognize the augmentation difference between the original graph and the augmented graph, and (iii) a third loss function defined to make the prediction results for the original graph and the augmented graph the same.
8. A memory for storing a graph classification model including an encoder that outputs a representation of a graph and a classifier connected to the encoder that predicts the class of the graph based on the representation of the graph; and It includes a control unit that classifies one or more graphs by inputting one or more graphs into the graph classification model above, and The above graph classification model is, A graph classification device that is trained to recognize an augmentation difference between the original graph and the augmented graph using the output of a graph representation learning model trained to approximate the distance between the original graph and the augmented graph by taking the representation of the original graph and the representation of the augmented graph as inputs, and the distance.
9. A computer-readable recording medium having a program for performing the method described in paragraph 1.
10. A computer program that is performed by a graph classification device and is stored on a recording medium to perform the method described in claim 1.

Description

Graph Classification Method and Apparatus Based on Augmentation-Aware Representation Learning The embodiments disclosed in this specification relate to a graph classification method and apparatus based on augmented recognition representation learning, and more specifically, to a method and apparatus for learning a graph classification model by recognizing differences in augmented graphs and accurately classifying graphs through the learned graph classification model. Graph classification is treated as an important problem in various application fields, such as social network analysis, web data mining, compound discovery, and molecular property prediction. Graph Neural Networks (GNNs) are used as one method to classify multiple graphs into classes. Graph neural networks are deep learning models for processing graph data that derive representations of vertices and graphs using algorithms that transmit messages between vertices. Graph neural networks have demonstrated better performance than existing graph kernel-based methods due to their ability to learn high-dimensional information about graph structures. However, simply using graph neural networks leads to overfitting problems due to reasons such as an insufficient number of graphs. Graph augmentation techniques were introduced to prevent overfitting in such graph neural networks, but simply using graph augmentation limits the representational power of the graph neural network. In other words, there is a problem where classification becomes inaccurate. In particular, existing graph augmentation techniques such as Node Drop, Edge Drop, and Shuffling cause significant variability among augmented graphs. These techniques have a problem in that they fail to accurately measure actual differences between graphs because they make the incorrect assumption that the same augmentation rate implies the same difference between graphs. Therefore, a graph classification technique based on a graph neural network that considers the differences in augmented graphs is required. For reference, Patent Document 1 is an invention relating to a graph classification method through comparative curriculum learning and a graph classification device for performing the same. Patent Document 1 discloses comparative learning content using augmented graphs, but does not provide a graph classification technology that takes into account the difference of augmented graphs. The attached drawings illustrative of preferred embodiments disclosed in this specification serve to further enhance understanding of the technical concept disclosed in this specification, along with specific details for implementing the invention; therefore, the contents disclosed in this specification should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is a block diagram illustrating the functional configuration of a graph classification device according to one embodiment. FIG. 2 is a flowchart illustrating the operation of a graph classification device learning a graph classification model according to one embodiment. FIG. 3 is a diagram conceptually illustrating the overall operation of a graph classification device according to one embodiment. FIG. 4 is a diagram illustrating the relationship between a graph classification model and a graph representation learning model processed by a graph classification device according to one embodiment. FIG. 5 is a diagram illustrating the augmented recognition learning operation and graph distance-based difference calculation operation of a graph classification device according to one embodiment. FIG. 6 is a diagram illustrating a graph of compounds processed by a graph classification device according to one embodiment. FIG. 7 is a diagram illustrating the consistency normalization operation of a graph classification device according to one embodiment. FIG. 8 is a flowchart illustrating a graph classification method according to one embodiment. FIGS. 9 and FIGS. 10 are diagrams illustrating graph classification performance simulated according to embodiments. Various embodiments are described in detail below with reference to the attached drawings. The embodiments described below may be implemented in various different forms. In order to explain the features of the embodiments more clearly, detailed descriptions of matters widely known to those skilled in the art to which the following embodiments belong have been omitted. Additionally, parts of the drawings unrelated to the description of the embodiments have been omitted, and similar parts throughout the specification have been given similar reference numerals. Throughout the specification, when a configuration is described as being "connected" to another configuration, this includes not only cases where they are "directly connected," but also cases where they are "connected with another configuration in between." Furthermore, when a configuration is described as "including" another