KR102817872B1 - Apparatus and method for predicting feature of node - Google Patents

Abstract

A device and method for predicting characteristics of a vertex are disclosed. The device for predicting characteristics of a vertex comprises: an input/output unit for receiving data and outputting a result of processing the data; a storage unit for storing a program for performing a method for predicting characteristics of a vertex; and a control unit for predicting characteristics of a vertex included in data received through the input/output unit by executing the program, wherein the control unit is characterized in that it generates a latent variable using a preset probability model having a multivariate Gaussian distribution for each vertex included in the data, and models the generated latent variable.

Description

{APPARATUS AND METHOD FOR PREDICTING FEATURE OF NODE}

Embodiments disclosed in this specification relate to a device and method for predicting characteristics of a vertex, and more particularly, to a device and method for predicting characteristics of a vertex included in data expressed in a graph form.

Recently, interest in graph data is increasing day by day. In such graph data, there are entities (entities or nodes) representing users or products, and the entities have feature vectors. At this time, the feature vectors can be used as key information for understanding the entities and the entire graph data. Meanwhile, graph data is observed in various ways in the real world. For example, social network services such as Facebook and Twitter, streaming services such as Wave and Netflix, and online shopping services such as Coupang and Gmarket use graph data based on relationships between users (or products).

With the growing interest in graph data, graph neural networks (GNN) models are being developed that map each entity to a low-dimensional space by combining and regenerating the feature vectors of each entity using the graph structure.

At this time, the above-described graph neural network model operates under the assumption that all entities in the graph data have feature vectors, but it is difficult to obtain features for all entities in real-world data graphs. For example, users of online social networks set their profiles to private, and e-commerce sellers often register items without useful descriptions, so that only some entities often have feature vectors. Therefore, there is a problem that performance may be degraded when performing tasks (e.g., vertex classification, etc.) using entity features of graph data.

Accordingly, techniques for predicting the characteristics of entities are being studied, but conventional techniques for predicting the characteristics of entities have encountered the problem of causing overfitting.

Korean Patent Publication No. 10-2018-007717 (Published on July 9, 2018)

The embodiments disclosed in this specification aim to provide a device and method for predicting the characteristics of a vertex, which can solve the overfitting problem by modeling the prior distribution of latent variables as a Gaussian Markov Random Field (GMRF) and performing regularization to force the latent variables to be correlated with each other according to the structure of the given graph data.

Other objects and advantages of the present invention can be understood by the following description, and will be more clearly understood by an embodiment. In addition, it will be easily understood that the objects and advantages of the present invention can be realized by the means and combinations thereof indicated in the claims.

As a technical means for achieving the above-described technical task, a device for predicting the characteristics of a vertex comprises: an input/output unit for receiving data and outputting a result of processing the data; a storage unit for storing a program for performing a method for predicting the characteristics of a vertex; and a control unit for predicting the characteristics of a vertex included in data received through the input/output unit by executing the program, wherein the control unit is characterized in that it generates a latent variable using a preset probability model having a multivariate Gaussian distribution for each vertex included in the data, and models the generated latent variable.

According to another embodiment, a method for predicting a characteristic of a vertex performed by a device for predicting a characteristic of a vertex includes the steps of: generating a latent variable using a preset probability model for each vertex included in data; and modeling the generated latent variable.

According to another embodiment, the recording medium is a computer-readable recording medium having recorded thereon a program for performing a method for predicting characteristics of a vertex. The method for predicting characteristics of a vertex, performed by a device for predicting characteristics of a vertex, includes: a step of generating a latent variable using a preset probability model for each vertex included in data; and a step of modeling the generated latent variable.

According to another embodiment, a computer program is a computer program stored in a recording medium for performing a method of predicting the characteristics of a vertex, which is performed by a device for predicting characteristics of a vertex. The method of predicting characteristics of a vertex, performed by the device for predicting characteristics of a vertex, includes a step of generating a latent variable using a preset probability model for each vertex included in data; and a step of modeling the generated latent variable.

According to one of the aforementioned problem solving means, the overfitting problem is solved by modeling the prior distribution of latent variables as a Gaussian Markov Random Field (GMRF) and performing regularization to force the latent variables to be correlated with each other according to the structure of the given graph data, thereby enabling more accurate prediction of the characteristics of the vertices.

According to another of the aforementioned problem solving methods, there is an effect of improving the performance of node classification by more accurately predicting the characteristics of vertices included in data expressed in a graph form.

The effects that can be obtained from the disclosed embodiments are not limited to the effects mentioned above, and other effects that are not mentioned will be clearly understood by those skilled in the art to which the disclosed embodiments belong from the description below.

The attached drawings below illustrate preferred embodiments disclosed in this specification, and serve to further understand the technical ideas disclosed in this specification together with specific contents for carrying out the invention, so the contents disclosed in this specification should not be interpreted as being limited to matters described in such drawings.
FIG. 1 is a functional block diagram of a vertex characteristic prediction device according to one embodiment.
FIG. 2 is a diagram for explaining a process in which a vertex characteristic prediction device according to one embodiment predicts a vertex characteristic.
FIG. 3 is a diagram for explaining a Gaussian Markov Random Field (GMRF) according to one embodiment.
FIG. 4 is a flowchart for explaining a method for predicting a characteristic of a vertex in a vertex characteristic prediction device according to one embodiment.

Below, various embodiments are described in detail with reference to the attached drawings. The embodiments described below may be modified and implemented in various different forms. In order to more clearly explain the features of the embodiments, detailed descriptions of matters that are widely known to those skilled in the art to which the embodiments belong are omitted. In addition, parts in the drawings that are not related to the description of the embodiments are omitted, and similar parts are given similar drawing reference numerals throughout the specification.

Throughout the specification, when a component is said to be "connected" to another component, this includes not only the case where it is "directly connected" but also the case where it is "connected with another component in between." Furthermore, when a component is said to "include" another component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated.

Hereinafter, embodiments will be described in detail with reference to the attached drawings.

But before explaining this, let's first define the meaning of the terms used below.

A graph is simply a data structure that gathers nodes and edges connecting those vertices. In other words, a graph is a term referring to a computer data structure, and it is a data structure that can express relationships between connected objects, and it is data composed of vertices and edges. In other words, a graph is data composed of multiple nodes and edges connecting the vertices, and each vertex has a feature vector that represents the characteristics of the corresponding vertex. Data in the form of a graph can be observed in various ways in the real world, and for example, relationships between users can be expressed as graph data in social network services such as Facebook or Twitter, streaming services such as Wave or Netflix, and shopping sites such as Coupang or Gmarket. In addition, data in the form of a graph can be expressed in the field of chemistry that identifies and classifies the characteristics of compounds such as drugs or proteins. In data expressed in the form of a graph, vertices represent entities (nodes) in the real world, and edges connecting them represent relationships between entities (nodes). Accordingly, in explaining the embodiments, data may mean data expressed in the form of a graph.

In addition to the terms defined above, terms that require explanation are explained separately below.

FIG. 1 is a functional block diagram of a vertex characteristic prediction device according to one embodiment, FIG. 2 is a diagram for explaining a process by which a vertex characteristic prediction device according to one embodiment predicts a vertex characteristic, and FIG. 3 is a diagram for explaining a Gaussian Markov random field (GMRF) according to one embodiment.

Referring to FIG. 1, a device (100) for predicting characteristics of a vertex according to one embodiment includes an input/output unit (110), a storage unit (120), and a control unit (130).

The input/output unit (110) may include an input unit for receiving input from a user and an output unit for displaying information such as the result of performing a task or the status of the vertex characteristic prediction device (100). That is, the input/output unit (110) is configured to receive data and output the result of processing the data. The vertex characteristic prediction device (100) according to the embodiment may receive data in the form of a graph through the input/output unit (110).

The storage unit (120) is a configuration in which files and programs can be stored, and can be configured through various types of memory. In particular, the storage unit (120) can store data and programs that enable the control unit (130) described below to perform operations for predicting the characteristics of a vertex according to the algorithm presented below.

The control unit (130) is a configuration including at least one processor, such as a CPU, Arduino, etc., and can control the overall operation of the vertex characteristic prediction device (100). That is, the control unit (130) can control other configurations included in the vertex characteristic prediction device (100) to perform an operation for predicting the vertex characteristic. The control unit (130) can perform an operation for predicting the vertex characteristic according to the algorithm presented below by executing a program stored in the storage unit (120). The method by which the control unit (130) performs an operation for predicting the vertex characteristic will be described later.

Below, a detailed description will be given of a process for predicting the characteristics of a vertex according to one embodiment by having the control unit (130) execute a program stored in the storage unit (120).

The control unit (130) can more accurately predict the characteristics of a node by using a Markov Graph Autoencoder (MGA) described below.

Meanwhile, a Markov graph autoencoder (MGA) can be composed of a network including an encoder and a decoder, where the encoder generates latent variables for each vertex, and the decoder can generate a high-dimensional feature vector from the latent variables. Meanwhile, the latent variables can include features for predicting the characteristics of the vertices. At this time, the Markov graph autoencoder (MGA) can be composed of one encoder (as shown in Fig. 2). )(210) and two decoders performing different roles ( and )(220). Meanwhile, the two decoders (220) may include a feature decoder and a label decoder. A description related to the Markov Graph Autoencoder (MGA) will be described later.

The control unit (130) can generate a unique one-hot feature vector for each vertex included in the data. At this time, the one-hot feature vector can be one of the methods for assigning initial features to vertices without features. However, the partial imputation as described above can cause imbalance between vertices depending on whether features are assigned, thereby increasing the risk of overfitting. Therefore, normalization may be necessary to solve this overfitting problem. A more detailed description of normalization will be provided later.

The control unit (130) generates a latent variable using a preset probability model for each vertex included in the data. More specifically, the control unit (130) can generate a latent variable for each vertex included in the data using a Gaussian Markov random field (GMRF) using an encoder. At this time, the Gaussian Markov random field (GMRF) may be a probability model used for normalization, and more specifically, it is described as follows.

<Gaussian Markov Random Field (GMRF)>

A Gaussian Markov Random Field (GMRF) is a graphical model representing a multivariate Gaussian distribution. Data (where vertices have continuous signals that are correlated in a graph structure) ), a Gaussian Markov random field (GMRF) is given for all vertices ( ) and main line ( ) for two potential functions ( and ) represents a signal distribution with a possible value ( ) with a random variable ( ) is assumed.

Specifically, each vertex ( ) for the vertex potential function ( ) and each edge ( ) for the edge potential function ( ) are defined as mathematical expressions 1 and 2 below, respectively.

Here, class is a parameter of the Gaussian Markov Random Field (GMRF), is the number of vertices. The nonzero elements of correspond to the edges of the data, as shown in Fig. 3. In Fig. 3, the Gaussian Markov Random Field (GMRF) is a parameter and Gaussian distribution . Here, The non-zero elements of the data ( ) corresponds to the edge of the joint probability ( ) is composed of the product of all potential functions. The joint probability can be expressed as the mathematical expression 3 below.

Here, is a normalization constant. Each potential function is or A parameter and The probability that the current probabilistic assumption will appear using is measured, and the joint probability is calculated by multiplying the potential function for all vertices and edges. If actual graph-shaped data is represented as a Gaussian Markov Random Field (GMRF), the relationship between vertices can be probabilistically understood. That is, in the Gaussian Markov Random Field (GMRF), a multivariate Gaussian distribution can be created that models the correlation between variables that follow the graph structure.

Gaussian Markov Random Field (GMRF) can be a probabilistic model that integrates real data in a probabilistic framework. Parameters and Their role can be understood in relation to the distributions they represent. is the covariance ( ) is the inverse of If this is small, a pair of signals ( and ) is likely to be observed. Is In the general case, it is desirable to set it to 0, since it determines the average of the signal when it is fixed and for simplicity of calculation, it is assumed that there is no initial bias of the signal.

In other words, the control unit (130) can perform normalization by modeling the prior distribution of the latent variables using a Gaussian Markov Random Field (GMRF) so that the distribution of the latent variables approaches the prior distribution by using an encoder, so that the latent variables have correlations with each other according to the structure of the given graph data. Here, the normalization is to model the prior distribution of the latent variables using a Gaussian Markov Random Field (GMRF) so that the distribution of the latent variables approaches the prior distribution, so that the latent variables have correlations with each other according to the structure of the given graph. In other words, the control unit (130) can generate latent variables by assuming the prior variables of the latent variables as Gaussian Markov Random Fields (GMRF) by using the graph structure to model the correlations between variables in a probabilistic framework. As described above, the control unit (130) can prevent the overfitting problem from occurring by modeling the prior distribution of the latent variables as a Gaussian Markov random field (GMRF) and performing normalization so that the latent variables are correlated with each other according to the structure of the given graph.

The control unit (130) models the generated latent variables. That is, the control unit (130) can predict the characteristics of each vertex based on the generated latent variables. At this time, predicting the characteristics of the vertex may be predicting a high-dimensional characteristic vector for each vertex by utilizing the generated latent variables.

More specifically, the control unit (130) can generate a feature vector and a label of a vertex by executing mutation inference on graph-structured data using a feature decoder and a label decoder to model the latent variable. At this time, the control unit (130) can maximize the likelihood of the feature and label of a given vertex by executing mutation inference to model the latent variable. At this time, the mutation inference can be for joint learning, and more specifically, it is described as follows.

<Variational inference>

The present invention relates to data ( ) Given an adjacency matrix A, the observed characteristics Wow label Likelihood of The optimal parameters that maximize is to find each vertex ( ) for latent variables , and realization of all latent variables. is expressed as . Here, is the number of vertices, is the size of the variable. Latent variable is each vertex ( ) characteristics ( ) includes features to predict the probability. However, the probability = The direct maximization of for It includes the difficult expectation of handling. Therefore, according to the embodiment, the evidence lower bound (ELBO) is changed to maximize the mutational inference. The equation for maximizing the likelihood can be expressed as the following mathematical expression 4.

Here, is the lower bound of evidence (ELBO), Is Parameterized distribution of and, Is and is a parameterized distribution of each latent variable ( ) is a characteristic ( ) and labels ( ) enough vertices to generate ) is expected to have information on the given go , , Assuming conditional independence between . Then, in the mathematical expression 4 described above, The first term of can be re-expressed as in mathematical equation 5 below.

Here, and is a set of vertices where features and labels are observed, respectively. Is It represents .

According to the mathematical expression 5 described above, Given, it can represent the conditional likelihood of the observed features and labels. Therefore, maximizing Equation 5 is equivalent to minimizing the reconstruction error of the observed variables in a general autoencoder. On the other hand, Equation 4 described above The KL divergence term is the distribution of the latent variable ( ) is a prior distribution ( ) can act as a regularizer to make it closer to the prior distribution ( ) which plays an essential role in the framework. Meanwhile, the characteristic of regularization is that the prior distribution ( ) may vary depending on how the model is modeled.

In other words, the control unit (130) can maximize the likelihood for given features and labels by modeling latent variables by executing mutational inference as described above. That is, the control unit (130) can perform mutational inference, but by changing the objective function to an evidence lower bound (ELBO) to perform learning, thereby restricting the distribution of latent variables, so that a large amount of calculation (exponential calculation) is not required in calculating the likelihood maximization formula, and thus the calculation can be performed in a shorter time.

<Prior distribution normalization>

According to the present invention, in order to consider the correlation between variables in probability modeling, a prior distribution ( ) is modeled as a Gaussian Markov Random Field (GMRF). At this time, the following parameterization can be applied to model the distribution of latent variables.

1. Basic parameterization

According to the embodiment, is a multivariate Gaussian distribution ( ) is modeled as, where, and are each different sizes and is the covariance matrix of all vertices. The elements are assumed to share the same covariance matrix. and is a set of learnable parameters ( ) containing the encoder function ( and ) are generated respectively. After that, the parameters and Using a prior distribution ( ) as a Gaussian Markov Random Field (GMRF). It is modeled as a graph Laplacian matrix using symmetric regularization. Information matrix from Makes.

<Information matrix>

Here, is the identity matrix, Is is a degree matrix. The result is data( ) is maintained as a positive semi-definite matrix that satisfies the constraints of the Gaussian Markov Random Field (GMRF). Except for the diagonal. Non-zero items of It corresponds to the item of . is a constant, since it represents a fixed prior distribution.

and Given a Gaussian model of , Divergence can be expressed as mathematical expression 6 below.

Here, Is and is a constant related to .

The computational bottleneck of the mathematical expression 6 described above is , and the calculation is Therefore, we have the covariance and rectangular matrix It is decomposed into . Here, and silver are hyperparameters. Meanwhile, log can be expressed as mathematical formula 7 below.

Here, class are each class is the unit matrix of size. The calculation of mathematical expression 7 is , which is the entire covariance matrix. It is more efficient.

For each inference, respectively and Generated from and Based on at Randomly samples the gradient based update. Since direct sampling of the , is not possible, we use the reparametrization trick of the variational autoencoder. This can be expressed as the following mathematical expression (8).

Here, and is a matrix of standard normal variables, supporting backpropagation. Each time, the sample is randomly sampled to simulate the sampling of .

2. Integrated deterministic modeling

Using the reparametrization trick, each inference time is different. can be approximated by sampling the expected term of the above-described mathematical expression 5. However, considering the characteristics of the problem that inference must be performed for all vertices at once rather than independently for each vertex and that meaningful observations must be made only for a part of the target variables, the sampling process can make the training unstable. Therefore, according to the present invention, the above-described basic parameterization can be improved by applying the following contents.

first, Dirac's delta function Change to . Here, 0 is It is a zero matrix of size . This is, The sampling process in all inferences The stability of training can be improved by replacing the two parameter matrices with a deterministic function that returns and to a single matrix Integrated with and integrated encoder It is generated in . This is and Normalization for all is a single matrix Because it is applied to deterministic modeling as well, The normalization effect of the divergence term can be maintained. In this way, by the integrated deterministic modeling method described above, Assuming that, the mathematical expression 6 described above The first two terms become equivalent.

Meanwhile, given the integrated deterministic modeling and the theorem below, the mathematical expression 6 described above Divergence can be changed into a general regularization function, which can be expressed as Equation 9 below.

<Summary>

Here, Is is a constant that is independent of .

Here, here, is the mathematical formula 6 is a constant such as .

The first term in Equation 9 is called the graph Laplacian regularizer, which is widely used in graph learning.

Meanwhile, in mathematical formula 9 can be expressed as mathematical formula 10 below.

Here, and Is The second and It is the th row, and are each vertex and is the degree of . The minimization of Equation 10 is the degree of the adjacent vertices. To have similar expressions in , The symmetric regularization of alleviates the influence of different vertex degrees in the regularization. Meanwhile, the second term of the above-mentioned mathematical expression 9 is can be thought of as measuring the amount of space that an object occupies. That is, if we maximize are distributed sparsely, compressing the embedding into a small space. can alleviate the influence of hyperparameters. Meanwhile, controls the balance between two terms with opposing goals.

Referring to FIG. 2, the control unit (130) according to the present invention uses a Markov graph autoencoder (MGA) to distribute the target by a deep neural network. , and can be expressed. A Markov graph autoencoder (MGA) according to one embodiment can form a network structure including an encoder (210) and a decoder (220), and a more detailed description thereof is as follows.

1. Encoder

Encoder( ) is a parameter ( ) using can be expressed. According to an embodiment, the encoder ( ) can be used to model the graph convolutional network (GCN). However, it is not limited to this and various graph neural networks can be used instead. Encoder ( ) output is distributed ( ) parameters , which is used as a latent variable due to the integrated deterministic modeling described above. becomes. When using a graph convolutional network (GCN) with two layers, the encoder ( )Is is defined as, where, is a normalized adjacency matrix, Is , is the activation function, is the vertex feature matrix, Is The order matrix is the weight matrix for the layer. On the other hand, when using an encoder based on a graph convolutional network (GCN), since we only have partial observations of the vertex features, the input matrix There is a problem that the initial feature can be assigned to the vertices without features, such as generating a one-hot feature vector. However, this partial imputation can cause imbalance between vertices depending on whether or not the feature is assigned, which can increase the risk of overfitting. Therefore, according to the embodiment, the identity matrix second By adopting the encoder ( ) can delete all observations from the input. This is Weight matrix regardless of whether of Each vertex in the th row ( ) for the encoder( ) is an independent embedding vector can be learned. Meanwhile, the weight matrix Introducing the identity feature matrix may cause limitations, since it may make the training process unstable due to the large number of parameters in the encoder. Therefore, ) generated vertex representation The vertex ( ) for each vector Normalized to unit hypersphere, it does not change the main function of creating a diverse representation of vertices that provides sufficient information for generating high-dimensional features, but it can improve the stability of training by restricting the output space.

2. Decoder

According to an embodiment, two decoders and Using each and Modeling. At this time, latent variables It is assumed that the observed features and labels have sufficient information to construct them. According to this, The complexity of the decoder can be minimized by choosing the simplest linear transformation, such as: , , and are learnable weights and biases, The number of silver characteristics, is the number of classes. Distribution using the output of the decoder. and How to model can vary depending on the properties of the data set. Many graph data sets for vertex classification have binary features and a single label for each vertex. Therefore, Assuming a multivariate Bernoulli distribution, is assumed to be a categorical distribution. Accordingly, loss terms can be generated as shown in Equations 11 and 12.

Here, and is the output of the decoder, is a logistic sigmoid function, which returns 1 if the condition holds, and 0 otherwise.

According to the present embodiment, one encoder and two decoders can be updated in an end-to-end manner using a gradient-based optimizer, and the overall objective function can be expressed as shown in the following mathematical expression 13 to minimize the loss condition.

Here, is a regularizer as shown in the mathematical expression 9 described above.

Meanwhile, two things can be changed in the proof lower bound (ELBO) of Equation 4. First, the hyperparameter Adjust the amount of regularization using . After that, binary cross entropy (BCE) loss A modified version of this can be used to balance the 0 and non-0 items of the real feature by adjusting the class weights according to occurrence.

According to the above-described present invention, by modeling the prior distribution of latent variables as a Gaussian Markov Random Field (GMRF) and performing regularization to force the latent variables to be correlated with each other according to the structure of the given graph data, the overfitting problem is solved, thereby enabling more accurate prediction of the characteristics of the vertices. In addition, there is an effect of improving the performance of the vertex classification by more accurately predicting the characteristics of the vertices included in the data expressed in the form of a graph.

FIG. 4 is a flowchart illustrating a method for predicting characteristics of a vertex in a device for predicting characteristics of a data vertex according to one embodiment.

The method for predicting the characteristics of a vertex according to the embodiment illustrated in FIG. 4 includes steps that are processed in time series in the device (100) for predicting the characteristics of a vertex illustrated in FIGS. 1 to 3. Therefore, even if the content is omitted below, the content described above with respect to the device (100) for predicting the characteristics of a vertex illustrated in FIGS. 1 to 3 can also be applied to the method for predicting the characteristics of a vertex according to the embodiment illustrated in FIG. 4.

Referring to FIG. 4, in step S410, a device (100) for predicting the characteristics of a vertex generates a latent variable using a preset probability model for each vertex included in the data. At this time, the preset probability model may be a Gaussian Markov Random Field (GMRF). Meanwhile, the device (100) for predicting the characteristics of a vertex may include a Markov graph autoencoder (MGA) configured as a network including an encoder and a decoder. At this time, the encoder may generate latent variables for each vertex, and the decoder may generate a high-dimensional feature vector from the latent variables. The device (100) for predicting the characteristics of a vertex may generate a latent variable for each vertex included in the data using a Gaussian Markov Random Field (GMRF) by using an encoder.

In step S420, the device (100) for predicting the characteristics of a vertex models the latent variables generated in step S410. The device (100) for predicting the characteristics of a vertex can predict the characteristics of each vertex based on the generated latent variables. At this time, predicting the characteristics of a vertex may be predicting a high-dimensional feature vector for each vertex by utilizing the generated latent variables. More specifically, the device (100) for predicting the characteristics of a vertex can generate the feature vector and label of a vertex by executing mutation inference on graph-structured data using a feature decoder and a label decoder to model the latent variables. At this time, the device (100) for predicting the characteristics of a vertex can maximize the likelihood of the characteristics and labels of a given vertex by executing mutation inference to model the latent variables.

The term '~ unit' used in the above embodiments means a software or a hardware component such as an FPGA (field programmable gate array) or an ASIC, and the '~ unit' performs certain roles. However, the '~ unit' is not limited to software or hardware. The '~ unit' may be configured to be in an addressable storage medium and may be configured to reproduce one or more processors. Thus, as an example, the '~ unit' includes components such as software components, object-oriented software components, class components, and task components, and processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables.

The functionality provided within the components and '~subunits' may be combined into a smaller number of components and '~subunits' or separated into additional components and '~subunits'.

Additionally, the components and '~parts' may be implemented to trigger one or more CPUs within the device or secure multimedia card.

Meanwhile, the method for predicting the characteristics of a vertex according to an embodiment described through this specification may also be implemented in the form of a computer-readable medium that stores instructions and data executable by a computer. At this time, the instructions and data may be stored in the form of program codes, and when executed by a processor, may generate a predetermined program module to perform a predetermined operation. In addition, the computer-readable medium may be any available medium that may be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable medium may be a computer recording medium, and the computer recording medium may include both volatile and nonvolatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. For example, the computer recording medium may be a magnetic storage medium such as an HDD and an SSD, an optical recording medium such as a CD, a DVD, and a Blu-ray disc, or a memory included in a server accessible through a network.

In addition, the method for predicting the characteristics of a vertex according to an embodiment described through this specification may be implemented as a computer program (or a computer program product) including instructions executable by a computer. The computer program includes programmable machine instructions processed by a processor, and may be implemented in a high-level programming language, an object-oriented programming language, an assembly language, a machine language, etc. In addition, the computer program may be recorded on a tangible computer-readable recording medium (e.g., a memory, a hard disk, a magnetic/optical medium, or a solid-state drive (SSD)).

Accordingly, the method for predicting the characteristics of a vertex according to an embodiment described through this specification can be implemented by executing the computer program as described above by a computing device. The computing device can include at least some of a processor, a memory, a storage device, a high-speed interface connecting the memory and a high-speed expansion port, and a low-speed interface connecting the low-speed bus and the storage device. Each of these components is connected to each other using various buses, and can be mounted on a common motherboard or mounted in another suitable manner.

Here, the processor may process instructions within the computing device, such as instructions stored in a memory or storage device for displaying graphical information for providing a GUI (Graphical User Interface) on an external input/output device, such as a display connected to a high-speed interface. In other embodiments, multiple processors and/or multiple buses may be utilized, as appropriate, together with multiple memories and memory types. Additionally, the processor may be implemented as a chipset comprising multiple independent analog and/or digital processors.

Additionally, memory stores information within a computing device. For example, memory may be comprised of volatile memory units or a collection thereof. For another example, memory may be comprised of nonvolatile memory units or a collection thereof. Memory may also be another form of computer-readable media, such as, for example, a magnetic or optical disk.

And, the storage device can provide a large storage space to the computing device. The storage device can be a computer-readable medium or a configuration including such a medium, and can include, for example, devices within a storage area network (SAN) or other configurations, and can be a floppy disk device, a hard disk device, an optical disk device, a tape device, flash memory, or other semiconductor memory devices or arrays of devices.

The above-described embodiments are for illustrative purposes, and those skilled in the art will understand that the above-described embodiments can be easily modified into other specific forms without changing the technical idea or essential features of the above-described embodiments. Therefore, it should be understood that the above-described embodiments are exemplary and not restrictive in all respects. For example, each component described as a single entity may be implemented in a distributed manner, and likewise, components described as distributed may be implemented in a combined manner.

The scope of protection sought by this specification is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

100: A device for predicting the characteristics of a vertex
110 : Input/output section
120 : Storage
130 : Control Unit

Claims (12)

An input/output unit for receiving data, processing it, and outputting the results;
A storage unit in which a program for performing a method of predicting the characteristics of a vertex is stored;
A control unit including at least one processor, and predicting the characteristics of vertices included in data received through the input/output unit by executing the program; and
A network comprising an encoder and a feature decoder and a label decoder with different roles,
The above control unit,
A device for predicting characteristics of a vertex, characterized in that it predicts characteristics of a vertex, wherein a latent variable is generated for each vertex included in the data using a preset probability model having a multivariate Gaussian distribution, wherein the encoder is used to generate a latent variable for each vertex included in the data using a Gaussian Markov Random Field (GMRF), and the prior distribution of the latent variable is modeled using the Gaussian Markov Random Field (GMRF) so that the distribution of the latent variable approaches the prior distribution, and the latent variables are normalized so that they have a correlation with each other according to the structure of the given graph data, and the generated latent variable is modeled. delete delete In paragraph 1,
The above control unit,
Modeling the latent variables generated above,
A device for predicting the characteristics of a vertex, characterized in that it generates a characteristic vector and label of a vertex by modeling the latent variable by executing mutation inference on graph-structured data using the above-mentioned characteristic decoder and label decoder. In paragraph 1,
The above latent variables are,
A device for predicting characteristics of a vertex, characterized in that it includes features for predicting characteristics of a vertex. A method for predicting a feature of a vertex, wherein the device for predicting a feature of a vertex comprises a network including an encoder and a feature decoder and a label decoder having different roles,
A step of generating a latent variable using a preset probability model for each vertex included in the data; and
A step of modeling the generated latent variable; comprising:
The steps for generating the above latent variables are:
A step of generating a latent variable for each vertex included in the data using a Gaussian Markov Random Field (GMRF) using the above encoder; and
A method for predicting the characteristics of a vertex, comprising the step of modeling the prior distribution of the latent variable using the Gaussian Markov Random Field (GMRF) so that the distribution of the latent variable approaches the prior distribution by using the encoder, and normalizing the latent variables so that they are correlated with each other according to the structure of the given graph data. delete delete In paragraph 6,
The step of modeling the generated latent variable is:
A method for predicting a feature of a vertex, comprising the step of generating a feature vector and a label of the vertex by modeling the latent variable by performing variational inference on graph-structured data using a feature decoder and a label decoder. In paragraph 6,
The above latent variables are,
A method for predicting characteristics of a vertex, characterized in that it includes features for predicting characteristics of a vertex. A computer-readable recording medium having recorded thereon a program for performing the method described in Article 6. A computer program stored in a recording medium for performing the method described in claim 6, performed by a device for predicting the characteristics of a vertex.