Edge prediction method and device using accurate edge prediction model based on positive unlabeled data learning

Abstract

To provide a method and apparatus for predicting an edge adapted to correctly predict an edge having a probability to be coupled in an edge-incomplete graph structure.SOLUTION: A method for predicting an edge which is executed by an apparatus for predicting an edge includes a step for inputting an edge-incomplete graph to an edge predicting model to predict an edge having a probability to be coupled in an edge-incomplete graph structure. The edge predicting model is to process at least one edge observed in an edge-incomplete graph structure as positive data and at least one node pair uncoupled in the edge-incomplete graph structure as unlabeled data, thereby carrying out binary classification.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The embodiments disclosed herein relate to a method and apparatus for edge prediction, and more particularly to a method and apparatus for learning and utilizing an edge prediction model that accurately predicts edges with a probability of being connected in the future in an edge-incomplete graph.

[0002] This research was conducted as part of the "XVoice: Multimodal Speech Meta-Learning" project (IITP-2022-0-00641) of the Human-Centered Artificial Intelligence Innovation Source Technology Development Project of the Ministry of Science and ICT and the Institute for Information and Communications Technology Planning (IITP), the "(SW Starlab) Flexible and Efficient Model Compression Technology Development to Support Diverse Applications and Environments" project (IITP-2020-0-00894) of the SW Computing Industry Source Technology Development Project, the "Artificial Intelligence Graduate School Support (Seoul National University)" project (IITP-2021-0-01343) and the "Artificial Intelligence Innovation Hub Research and Development" project (IITP-2021-0-02068) of the Information and Communications Broadcasting Innovation Talent Development Project. [Background technology]

[0003] We can easily come across edge-incomplete graphs in our daily lives. Examples of edge-incomplete graphs include friendship relationships in social networks and citation relationships between papers. In social networks, users are nodes, and friendship relationships between users are considered edges. When adding friends on a social network, users do not need to verify all users, and relationships can include those in which users are actually friends but are not connected as friendship relationships on the social network. In a paper citation network, papers are nodes, and citation relationships between papers are edges. When a user cites a paper, they do not verify all published papers, so papers that should be cited may be missed.

[0004] Existing techniques for predicting edges in edge-incomplete graphs have the drawback of being heavily dependent on the graph in which a particular edge is incomplete. Existing techniques assume that all edges in a particular edge-incomplete graph are fully observed and do not consider unobserved missing links when learning. This makes it impossible to communicate information between nodes that can form edges but are not connected in the particular graph, leading to overfitting of the edge prediction model to the particular edge-incomplete graph.

[0005] Therefore, there is a need for an edge prediction technique that takes into account unconnected nodes in edge-incomplete graphs.

[0006] For reference, Patent Document 1 is an invention related to an apparatus and method for predicting node characteristics. However, Patent Document 1 only discloses the content for predicting graph nodes, and does not provide an edge prediction technology that takes into account nodes that are not connected in the graph. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Korean Patent Publication No. 10-2023-0083925 [Non-patent literature]

[0008] [Non-Patent Document 1] IITP-2022-0-00641 [Non-patent document 2] IITP-2020-0-00894 [Non-patent document 3] IITP-2021-0-01343 [Non-patent document 4] IITP-2021-0-02068 Summary of the Invention [Problem to be solved by the invention]

[0009] The embodiments disclosed herein aim to accurately predict edges that have a probability of being connected in an edge-incomplete graph structure using an edge prediction model that processes at least one observed edge in the edge-incomplete graph structure as positive data and at least one unconnected node pair in the edge-incomplete graph structure as unlabeled data.

[0010] Other objects and advantages of the present invention will become apparent from the following description and examples, and it will be readily apparent that the objects and advantages of the present invention can be realized by the means and combinations thereof disclosed in the claims. [Means for solving the problem]

[0011] As a technical means for achieving the above-mentioned technical object, an edge prediction method performed by an edge prediction device includes a step of inputting an edge incomplete graph to an edge prediction model and predicting edges having a probability of being connected in the structure of the edge incomplete graph, wherein the edge prediction model is a model that processes at least one observed edge in the structure of the edge incomplete graph as positive data and processes at least one unconnected node pair in the structure of the edge incomplete graph as unlabeled data to perform binary classification.

[0012] According to another embodiment, an edge prediction device includes a memory that stores an edge incomplete graph and an edge prediction model; and a control unit that inputs the edge incomplete graph to the edge prediction model and predicts edges that have a probability of being connected in the structure of the edge incomplete graph, wherein the edge prediction model is a model that performs binary classification by treating at least one observed edge in the structure of the edge incomplete graph as positive data and treating at least one unconnected node pair in the structure of the edge incomplete graph as unlabeled data.

[0013] According to yet another embodiment, the recording medium is a computer-readable recording medium having a program for executing the edge prediction method recorded thereon.

[0014] According to yet another embodiment, the computer program is a computer program stored on a recording medium for being executed by an edge prediction device and for performing an edge prediction method. [Effects of the Invention]

[0015] According to any one of the above-mentioned means for solving the problem, it is possible to provide an edge prediction method and device that can process edges observed in an edge incomplete graph structure as positive data and process unconnected node pairs in the edge incomplete graph structure as unlabeled data to classify the unconnected node pairs.

[0016] In addition, according to any one of the above-mentioned means for solving the problem, it is possible to provide an edge prediction method and device that enables information transmission between unconnected node pairs by using the structure of a predicted edge incomplete graph in which a random variable is applied between unconnected node pairs in an edge incomplete graph in the learning process of an edge prediction model.

[0017] In addition, according to any one of the above-mentioned problem-solving means, it is possible to provide an edge prediction method and device that can convert a predicted edge incomplete graph into a line graph and efficiently calculate the expected value for the random variables of the predicted edge incomplete graph by assuming it in a Markov network.

[0018] In addition, by using any one of the above-mentioned problem-solving means, it is possible to provide an edge prediction method and apparatus that can solve the oversmoothing problem by approximating the structure of the predicted edge incomplete graph and solve the training time that increases depending on the number of nodes.

[0019] In addition, according to any one of the above-mentioned problem solving means, it is possible to provide an edge prediction method and device that can efficiently approximate a predicted edge incomplete graph by selecting candidate edges based on the degree of nodes in the structure of the predicted edge incomplete graph, selecting edges with high weight values ​​based on random variables from the candidate edges, and removing the remaining edges.

[0020] In addition, according to any one of the above-mentioned means for solving the problem, it is possible to provide an edge prediction method and device that can gradually improve the quality of the edge prediction model and the quality of the predicted edge incomplete graph by repeating the process of updating the parameters of the edge prediction model using the random variables of the predicted edge incomplete graph and updating the random variables of the predicted edge incomplete graph using the predicted probability output by the edge prediction model.

[0021] In addition, according to any one of the above-mentioned solutions, it is possible to provide an edge prediction method and device that can improve the accuracy of an edge prediction model based on a binary loss function by taking into account newly added edges in a predicted edge incomplete graph and applying randomly sampled edges during the learning process of an edge prediction model, thereby balancing the number of connected edges and the number of unconnected edges in the structure of the edge incomplete graph.

[0022] In addition, according to any one of the above-mentioned solutions, it is possible to provide an edge prediction method and device that can prevent excessive self-reinforcement by taking into account newly added edges in a predicted edge incomplete graph during the learning process of an edge prediction model and by using a correction loss function that applies randomly sampled edges.

[0023] The effects obtained from the disclosed embodiments are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art to which the embodiments disclosed below pertain. [Brief explanation of the drawings]

[0024] The accompanying drawings below illustrate preferred embodiments disclosed in this specification and serve to facilitate a better understanding of the technical concepts disclosed in this specification, along with specific content for implementing the invention. Therefore, the content disclosed in this specification should not be interpreted as being limited solely to the matters depicted in the drawings.

[0025] [Figure 1] FIG. 2 is a block diagram illustrating a functional configuration of an edge prediction device according to an embodiment. [Figure 2] 10 is a diagram illustrating a graph transformed by an edge prediction device according to an embodiment; [Figure 3] 10 is a flowchart illustrating an operation of an edge prediction device according to an embodiment for learning an edge prediction model. [Figure 4] 1 is a flowchart illustrating an edge prediction method according to one embodiment. [Figure 5] FIG. 10 illustrates simulated edge prediction performance according to an embodiment. [Figure 6] FIG. 10 illustrates simulated edge prediction performance according to an embodiment. [Figure 7] FIG. 10 illustrates simulated edge prediction performance according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0026] Various embodiments will be described in detail below with reference to the accompanying drawings. The embodiments described below may be implemented in various modified forms. In order to more clearly describe the features of the embodiments, detailed descriptions of matters that are well known to those skilled in the art to which the following embodiments pertain will be omitted. In addition, parts of the drawings that are not relevant to the description of the embodiments will be omitted, and similar parts will be designated by similar reference numerals throughout the specification.

[0027] Throughout the specification, when a certain component is said to be "connected" to another component, this includes not only "directly connected" but also "connected via another component in between." Furthermore, when a certain component is said to "include" another component, this does not exclude the other component, but means that the other component may also be included, unless otherwise specified.

[0028] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings.

[0029] A "graph" is a data structure that collects nodes and edges (links) connecting those nodes. Each node has a feature vector that indicates the characteristics of that node. Graph-type data can be observed in various ways in the real world. For example, relationships between users on social networking services such as Facebook and Twitter, streaming services such as Wave and Netflix, and shopping sites such as Coupang and Gmarket can be expressed as graph-type data. Graph-type data can also be expressed in the field of chemistry, where the characteristics of compounds such as drugs and proteins are identified and classified. However, due to the practical limitations of being able to access all data in the real world, some edges may be missing. A graph with missing edges is called an edge-incomplete graph.

[0030] Table 1 below defines terms used in this specification.

[0031] [Table 1]

[0032] JPEG0007771287000002.jpg62159

[0033] This embodiment trains an edge prediction model based on positive-unlabeled data learning to accurately predict edges with a probability of being connected in an edge-incomplete graph. Positive-unlabeled data learning is a type of binary classification. While traditional binary classification aims to train a binary classification model using positive and negative training data instances, positive-unlabeled data learning aims to train a classification model using positive-unlabeled training data instances. That is, positive-unlabeled data learning aims to train a binary classification model when only a portion of the data is labeled as positive during the model training process and the rest is given unlabeled. Positive labeling refers to classifying the degree of agreement (identity or similarity) between the correct answer (actual value) and the predicted value, while negative labeling refers to classifying the degree of agreement (identity or similarity) between an incorrect value and the predicted value. Identity or similarity can be measured based on various distances between values. Positive unlabeled data learning does not provide training instances classified as negative during the model learning process, so existing binary classification models cannot be applied as is.

[0034] In this embodiment, observed edges of a specific edge-incomplete graph are considered as positive data instances, and the remaining unconnected node pairs (node ​​pairs that may be connected in the future) are treated as unlabeled data instances, which can be utilized when learning unconnected node pairs in a specific graph as connected or unconnected. In particular, this embodiment introduces a random variable indicating a connection relationship for unconnected node pairs, and then utilizes the expected value of the graph instead of a specific graph to learn an edge prediction model, thereby improving the accuracy and efficiency of positive unlabeled data learning.

[0035] FIG. 1 is a block diagram illustrating the functional configuration of an edge prediction device according to an embodiment.

[0036] Referring to FIG. 1, an edge prediction device 100 according to an embodiment may include an input / output unit 110, a memory 120, a control unit 130, and a communication unit 140.

[0037] 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 a result of execution of a task or a status of the edge prediction device 100. That is, the input / output unit 110 is configured to receive data and output a result of processing the data. The edge prediction device 100 according to the embodiment may receive an edge prediction request, etc., via the input / output unit 110.

[0038] The memory 120 is configured to store files and programs and may be configured with various types of memory. In particular, the memory 120 may store data and programs that enable the control unit 130 (described later) to perform calculations for edge prediction according to the algorithms presented below.

[0039] The memory 120 can store the edge incomplete graph and the edge prediction model. The memory 120 can store the predicted edge incomplete graph and the approximated predicted edge incomplete graph. The memory 120 can also store the predicted probabilities output by the edge prediction model.

[0040] The control unit 130 includes at least one processor such as a CPU, a GPU, or the like, and can control the overall operation of the edge prediction device 100. That is, the control unit 130 can control other components included in the edge prediction device 100 to perform an operation for edge prediction. The control unit 130 can execute a program stored in the memory 120 to perform a calculation for predicting edges from an edge incomplete graph according to an algorithm to be presented below.

[0041] The communication unit 140 may perform wired or wireless communication with other devices or networks. For example, if a server providing a service of a specific online platform that collects or processes data included in the graph is implemented as a separate device, the communication unit 140 may receive the edge-incomplete graph by communicating with the server providing the service of the online platform, and may provide the edge-incomplete graph in which edges are generated according to predicted probabilities or edges are completed based on the received edge-incomplete graph to the server or a user terminal.

[0042] For this purpose, the communication unit 140 may include a communication module supporting at least one of various wired and wireless communication methods, and the communication module may be implemented in the form of a chipset. The mobile communication or wireless communication supported by the communication unit 140 may be, for example, an N-generation mobile communication protocol, Wi-Fi (Wireless Fidelity), Wi-Fi Direct, Bluetooth, UWB (Ultra-Wide Band), or NFC (Near Field Communication).

[0043] The control unit 130 may input the edge incomplete graph to an edge prediction model to predict edges that have a probability of being connected in the structure of the edge incomplete graph. A model based on a graph convolution network may be applied to the edge prediction model, or other learning models capable of graph processing may also be applied.

[0044] The control unit 130 can perform a binary classification of the graph configuration and classify unconnected node pairs by using an edge prediction model to process edges observed in the edge-incomplete graph structure as positive data and to process unconnected node pairs in the edge-incomplete graph structure as unlabeled data.

[0045] The control unit 130 converts the edge-incomplete graph into another graph to improve the edge prediction accuracy and processing efficiency of the edge prediction model.

[0046] FIG. 2 is a diagram illustrating a graph converted by an edge prediction device according to an embodiment. 2, the control unit 130 can convert the edge incomplete graph into a predicted edge incomplete graph, where the predicted edge incomplete graph is a graph to which a random variable indicating the connection state of an unconnected node pair in the structure of the edge incomplete graph is applied.

[0047] The control unit 130 can convert two adjacent edges in the edge incomplete graph structure into a line graph represented by two connected nodes, and calculate the expected value of a random variable using a Markov network that models the joint probability distribution of the nodes of the converted line graph.

[0048] The control unit 130 can convert the structure of the predicted edge incomplete graph into an approximated predicted edge incomplete graph by setting the number of edges to be maintained in the structure of the predicted edge incomplete graph and not connecting the remaining node pairs except for node pairs whose connection probability is higher than a reference value.

[0049] In the process of learning the edge prediction model, the control unit 130 may propagate information in a graph convolution network of the edge prediction model using the predicted edge incomplete graph (or an approximated predicted edge incomplete graph) to learn the model.

[0050] During the process of learning the edge prediction model, the control unit 130 can update the parameters of the edge prediction model using a predicted edge incomplete graph (or an approximated predicted edge incomplete graph) to which a random variable indicating the connection state of unconnected node pairs in the structure of the edge incomplete graph is applied by the edge prediction model.

[0051] The control unit 130 may update the random variables of the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) using the predicted probability output by the edge prediction model during the process of learning the edge prediction model.

[0052] The control unit 130 may take into account the edges added in the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) during the process of learning the edge prediction model, and may learn the edge prediction model using a binary loss function that applies randomly sampled edges to balance between the number of connected edges and the number of unconnected edges in the structure of the edge incomplete graph.

[0053] The control unit 130 may take into account the edges added in the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) during the process of learning the edge prediction model, and may learn the edge prediction model using a correction loss function that prevents excessive self-reinforcement based on randomly sampled edges.

[0054] The control unit 130 may predict the edge using an edge prediction model optimized to minimize the binary loss function and the adjusted loss function.

[0055] FIG. 3 is a flowchart illustrating an operation of an edge prediction device according to an embodiment for learning an edge prediction model.

[0056] JPEG0007771287000003.jpg185170

[0057]

number

[0058]

number

[0059] JPEG0007771287000006.jpg194170

[0060] By repeating the process of updating the parameters of the edge prediction model using the random variables of the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) and updating the random variables of the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) using the predicted probabilities output by the edge prediction model, the quality of the edge prediction model and the quality of the predicted edge incomplete graph (or the approximated predicted edge incomplete graph) can be gradually improved.

[0061] The edge prediction device uses the edge prediction model f θ To learn (i, j), the sum of two loss functions is minimized. The binary loss function for positive data and unlabeled data is a loss function that balances the number of edges by sampling them to resolve imbalances in the actual graph, and can be expressed as in Equation 3. If the current parameters of the edge prediction model are inaccurate, the quality of the predicted edge-incomplete graph will deteriorate, leading to inaccurate parameters in the next iteration. To solve this problem, the corrected loss function, which measures the binary cross entropy for edges, can be expressed as in Equation 4. The sum of the binary loss function and the corrected loss function can be expressed as in Equation 5.

[0062]

number

[0063]

number

[0064]

number

[0065] The edge prediction device considers newly added edges in the predicted edge incomplete graph during the learning process of the edge prediction model and applies randomly sampled edges, thereby achieving a balance between the number of connected edges and the number of unconnected edges in the structure of the edge incomplete graph and improving the accuracy of the edge prediction model based on the binary loss function.

[0066] The edge prediction device takes into account newly added edges in the predicted edge incomplete graph during the learning process of the edge prediction model, and can improve the accuracy of the edge prediction model by preventing excessive self-reinforcement through a correction loss function that applies randomly sampled edges.

[0067] In step S350, the edge prediction device determines whether a learning termination condition is met. If the model converges or reaches the maximum number of iterations, model learning is interrupted. If the learning termination condition is not met, steps S320 to S340 are repeated. If the learning termination condition is met, the learned edge prediction model is output in step S360.

[0068] The algorithm for edge prediction described with reference to FIG. 3 can be expressed in pseudocode as shown in Table 2.

[0069] [Table 2]

[0070] The algorithm for such edge prediction is called PULL (PU-Learning-based Link predictor).

[0071] FIG. 4 is a flow chart illustrating an edge prediction method according to one embodiment.

[0072] The edge prediction method according to the embodiment shown in Fig. 4 includes steps that are processed in time series by the edge prediction device shown in Fig. 1 to Fig. 3. Therefore, even though the content will be omitted below, the content described above regarding the edge prediction device shown in Fig. 1 to Fig. 3 can also be applied to the edge prediction method according to the embodiment shown in Fig. 4.

[0073] Referring to FIG. 4, in step S410, the edge prediction device collects edge incomplete graphs.

[0074] In step S420, the edge prediction device inputs the edge incomplete graph into the edge prediction model and predicts edges having a probability of being connected in the structure of the edge incomplete graph.

[0075] The edge prediction model applied to the edge prediction method may be a model that performs binary classification by treating edges observed in the structure of an edge-incomplete graph as positive data and treating unconnected node pairs in the structure of an edge-incomplete graph as unlabeled data.

[0076] The edge prediction model applied to the edge prediction method may be a model in which the parameters of the edge prediction model are updated by using a predicted edge incomplete graph to which a random variable indicating the connection state of unconnected node pairs in the structure of the edge incomplete graph is applied.

[0077] The edge prediction model applied to the edge prediction method may be a model in which the expected value of a random variable is calculated by converting two adjacent edges in an edge incomplete graph structure into a line graph represented by two nodes connecting the two adjacent edges, and using a Markov network that models the joint probability distribution of the nodes of the converted line graph.

[0078] The edge prediction model applied to the edge prediction method may be a model that approximates the structure of the predicted edge incomplete graph by setting the number of edges to be maintained in the structure of the predicted edge incomplete graph and not connecting the remaining node pairs except for node pairs whose connection probability is higher than a reference value.

[0079] The edge prediction model applied in the edge prediction method may be a model that is learned by propagating information in a graph convolution network of the edge prediction model using a predicted edge incomplete graph.

[0080] The edge prediction model applied to the edge prediction method may be a model in which the random variables of the predicted edge incomplete graph are updated by using the predicted probabilities output by the edge prediction model.

[0081] The edge prediction model applied to the edge prediction method may be a model trained by a binary loss function that applies randomly sampled edges to balance the number of connected edges and the number of unconnected edges in the structure of the edge incomplete graph, taking into account the edges added in the predicted edge incomplete graph.

[0082] The edge prediction model applied to the edge prediction method may be a model trained by a corrective loss function that takes into account the added edges in the predicted edge-incomplete graph and prevents excessive self-reinforcement based on randomly sampled edges.

[0083] 5 to 7 are diagrams illustrating edge prediction performance simulated according to the embodiment.

[0084] As a result of comparing the performance of existing models with edge prediction problems using a total of five real-life graph data sets (PubMed, Cora-full, Chameleon, Crocodile, and Facebook), the PULL model of this embodiment showed the highest performance in terms of the performance indicators AUROC (Area Under ROC curve) and AUPRC (Area Under Precision-Recall Curve).

[0085] Referring to Figure 5, which shows the AUROC of PULL by iteration, the dotted line represents the actual number of edges, and it can be seen that PULL's performance improves as the iterations proceed. This indicates that PULL improves the quality of the predicted graph as the iterations progress, ultimately resulting in accurate predictions. For PubMed, Cora-full, and Chameleon, when the number of sampled edges k exceeds the actual number of edges, accuracy converges or slightly decreases. This is due to a smoothing problem that occurs when propagating information through a graph with more edges than the actual graph. For Crocodile and Facebook, prediction accuracy improves even when the number of sampled edges k is greater than the actual number of edges. This indicates that the actual graph structures of Crocodile and Facebook inherently contain leaky edges.

[0086] Referring to Figure 6, which shows the effect of the correction loss function on the edge prediction performance of PULL, PULL-L C is the corrected loss function L C indicates that PULL was not applied, and PULL-L C It can be seen that the predicted graph structure has consistently better performance than the actual graph. This is because when the predicted graph structure contains more edges than the actual graph, the corrected loss function L C This means that the performance degradation of PULL can be effectively prevented.

[0087] Referring to Figure 7, which shows the processing time of PULL for a sampled subgraph, we can see that the time increases linearly as the size increases, demonstrating the scalability of PULL to large graphs.

[0088] The term "module" used in the above embodiments refers to software or hardware components such as FPGAs (field programmable gate arrays) or ASICs, and the "module" performs a certain function. However, the term "module" is not limited to software or hardware. A "module" may be configured to reside on an addressable storage medium or to execute one or more processors. Thus, by way of example, "module" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables.

[0089] The functionality provided within components and units may be combined into fewer components and units or separated into additional components and units.

[0090] Furthermore, the components and "units" may be implemented to implement one or more CPUs within a device or a secure multimedia card.

[0091] Meanwhile, the edge prediction method according to an embodiment described herein may also be embodied in the form of a computer-readable medium storing computer-executable instructions and data. Here, the instructions and data may be stored in the form of program code, which, when executed by a processor, may generate a predetermined program module and perform a predetermined operation. Furthermore, the computer-readable medium may be any available medium accessible by a computer, including both volatile and nonvolatile media, and both separable and non-separable media. Furthermore, the computer-readable medium may be a computer recording medium. The computer recording medium may be any volatile and non-volatile, separable and non-separable medium 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 a hard disk drive (HDD) or solid-state drive (SSD), an optical storage medium such as a CD, DVD, or Blu-ray disc, or a memory included in a server accessible via a network.

[0092] Furthermore, the edge prediction method according to an embodiment described herein may be implemented as a computer program (or a computer program product) including computer-executable instructions. The computer program includes programmable machine instructions to be 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. Furthermore, the computer program may be recorded on a computer-readable recording medium (e.g., a memory, a hard disk, a magnetic / optical medium, or a solid-state drive (SSD)).

[0093] Therefore, the edge prediction method according to one embodiment described herein can be implemented by executing the above-described computer program on a computing device. The computing device may include a processor, a memory, a storage device, a high-speed interface connected to the memory and a high-speed expansion port, and at least a portion of a low-speed interface connected to a low-speed bus and the storage device. Each of these components is connected to each other using various buses and may be mounted on a common motherboard or in other suitable manners.

[0094] Here, the processor may process instructions within a computing device. Such instructions may include instructions stored in a memory or storage device for displaying graphic information to provide a GUI (Graphical User Interface) on an external input and output device, such as a display connected to a high-speed interface. In other embodiments, multiple processors and / or multiple buses may be used, along with multiple memories and memory types, as appropriate. Also, the processor may be implemented as a chipset consisting of chips containing multiple independent analog and / or digital processors.

[0095] Also, memory stores information within a computing device. As an example, memory may be comprised of a volatile memory unit or collection thereof. As another example, memory may be comprised of a non-volatile memory unit or collection thereof. Memory may also be in other forms of computer-readable media, such as magnetic or optical disks.

[0096] A storage device can provide a large amount of storage space to a computing device. A storage device can be a computer-readable medium or a configuration that includes such a medium, such as a device in a Storage Area Network (SAN) or other configuration, and can be a floppy disk drive, hard disk drive, optical disk drive, tape drive, flash memory, or other similar semiconductor memory device or device array.

[0097] The above-described embodiments are merely illustrative, and those skilled in the art will understand that the above-described embodiments may be easily modified into other specific forms without changing the technical ideas or essential features of the above-described embodiments. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and are not limiting. For example, each component described as a single component may be implemented in a distributed form, and similarly, each component described as a distributed component may be implemented in a combined form.

[0098] The scope of protection sought by this specification is determined by the claims set forth below rather than the above detailed description, and all modifications or variations derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included within the scope of the present invention. [Explanation of symbols]

[0099] 100 Edge Predictor 110 Input / output section 120 memory 130 control section 140 Communications Department

Claims

1. An edge prediction method executed by an edge prediction device, comprising: inputting an edge-incomplete graph into an edge prediction model and predicting edges having a probability of being connected in the structure of the edge-incomplete graph; The edge prediction method, wherein the edge prediction model is a model for performing binary classification by treating at least one observed edge in the structure of the edge incomplete graph as positive data and treating at least one unconnected node pair in the structure of the edge incomplete graph as unlabeled data.

2. The edge prediction method according to claim 1, wherein the edge prediction model is a model in which parameters of the edge prediction model are updated by using a predicted edge incomplete graph to which a random variable indicating the connection state of an unconnected node pair in the structure of the edge incomplete graph is applied.

3. 3. The edge prediction method according to claim 2, wherein the edge prediction model is a model in which the expected value of the random variable is calculated by converting the edge incomplete graph structure into a line graph represented by two nodes connecting two adjacent edges, and using a Markov network that models the joint probability distribution of the nodes of the converted line graph.

4. 3. The edge prediction method of claim 2, wherein the edge prediction model is a model that approximates the structure of the predicted edge incomplete graph by setting the number of edges to be maintained in the structure of the predicted edge incomplete graph and not connecting any remaining node pairs except for node pairs whose connection probability is higher than a reference value.

5. The edge prediction method according to claim 2 , wherein the edge prediction model is a model that is trained by propagating information in a graph convolution network of the edge prediction model using the predicted edge incomplete graph.

6. The edge prediction method according to claim 2 , wherein the edge prediction model is a model in which the random variables of the predicted edge incomplete graph are updated by using predicted probabilities output by the edge prediction model.

7. The edge prediction model is (i) a binary loss function applying randomly sampled edges to balance the number of connected and unconnected edges in the structure of the edge-incomplete graph, taking into account the added edges in the predicted edge-incomplete graph; and (ii) a corrective loss function that takes into account added edges in the predicted edge-incomplete graph to prevent excessive self-reinforcement based on the randomly sampled edges.

8. a memory for storing the edge incomplete graph and the edge prediction model; a control unit that inputs the edge incomplete graph into the edge prediction model to predict edges that have a probability of being connected in a structure of the edge incomplete graph, The edge prediction model is a model that performs binary classification by treating at least one observed edge in the structure of the edge incomplete graph as positive data and treating at least one unconnected node pair in the structure of the edge incomplete graph as unlabeled data.

9. A computer-readable recording medium having a program recorded thereon for executing the method of claim 1.

10. A computer program stored on a recording medium for executing the method of claim 1 when executed by an edge prediction device.

Images