KR-102961704-B1 - APPARATUS AND METHOD FOR PROCESSING SPATIOTEMPORAL DATA BASED ON GRAPH NEURAL CONTROLLED DIFFERENTIAL EQUATIONS

Abstract

The present invention relates to a spatiotemporal data processing apparatus and method based on a Graph Neural Controlled Differential Equation, wherein the apparatus comprises a preprocessing unit that generates a continuous path for each node in time series data; and a main processing unit that integrates and processes time information and spatial information by combining a Graph Convolution Network (GCN) and a Neural Controlled Differential Equation (NCDE) for the generated path, and the main processing unit can calculate a final hidden vector and predict an output layer by performing time processing and spatial processing for each node using two Controlled Differential Equation (CDE) functions.

Inventors

• 박노성
• 최정환
• 황지현
• 최황용

Assignees

• 연세대학교 산학협력단

Dates

Publication Date

20260507

Application Date

20221114

Claims (12)

1. A preprocessing unit that generates a continuous path for each node in time series data; and It includes a main processing unit that integrates and processes temporal and spatial information by combining Graph Convolution Networks (GCN) and Neural Controlled Differential Equation (NCDE) for the above-mentioned generated path, The above main processing unit A Graph Neural Controlled Differential Equation-based spatiotemporal data processing device characterized by performing temporal and spatial processing for each of the above nodes using two Controlled Differential Equation (CDE) functions to calculate the final hidden vector and predict the output layer.
2. In paragraph 1, the pretreatment unit A graph neural control differential equation-based spatiotemporal data processing device characterized by generating a continuous path by performing an interpolation algorithm for each node.
3. In paragraph 2, the above pretreatment unit A graph neural control differential equation-based spatiotemporal data processing device characterized by using a natural cubic spline as the interpolation algorithm.
4. In paragraph 1, the main processing unit A first NCDE module that generates a hidden trajectory of time information by temporally processing the continuous path of each of the above nodes; and A graph neural control differential equation-based spatiotemporal data processing device characterized by including a second NCDE module that generates a hidden trajectory of spatial information by spatially processing the continuous path of each of the above nodes.
5. In paragraph 4, the above-mentioned first NCDE module is A graph neural control differential equation-based spatiotemporal data processing device characterized by generating a matrix by stacking the above hidden trajectories for all nodes and converting each row of the matrix into a continuous RNN by individually processing the CDE function.
6. In paragraph 4, the main processing unit A graph neural control differential equation-based spatiotemporal data processing device characterized by further including an initial value generation module that generates initial values for the above-mentioned temporal processing and spatial processing, and trains the CDE function including parameters of the initial value generation layer, a node embedding matrix, and an output layer.
7. In a spatiotemporal data processing method performed in a spatiotemporal data processing device based on a Graph Neural Controlled Differential Equation, A preprocessing step that generates a continuous path for each node in time series data through a preprocessing unit; and It includes a main processing step that integrates and processes temporal and spatial information by combining Graph Convolution Networks (GCN) and Neural Controlled Differential Equation (NCDE) on the generated path through the main processing unit. The above main processing step is A graph neural controlled differential equation-based spatiotemporal data processing method characterized by performing temporal and spatial processing for each node using two CDE (Controlled Differential Equation) functions to calculate the final hidden vector and predict the output layer.
8. In claim 7, the above preprocessing step A graph neural control differential equation-based spatiotemporal data processing method characterized by generating a continuous path by performing an interpolation algorithm for each node.
9. In paragraph 8, the above preprocessing step A graph neural control differential equation-based spatiotemporal data processing method characterized by using a natural cubic spline as the interpolation algorithm.
10. In Clause 7, the above main processing step is A temporal processing step for generating a hidden trajectory of temporal information by temporally processing the continuous path of each node through the first NCDE module; and A graph neural control differential equation-based spatiotemporal data processing method characterized by including a spatial processing step that generates a hidden trajectory of spatial information by spatially processing the continuous path of each node through a second NCDE module.
11. In item 10, the above time processing step A spatiotemporal data processing method based on graph neural control differential equations, characterized by generating a matrix by stacking the above hidden trajectories for all nodes and converting each row of the matrix into a continuous RNN by individually processing the CDE function.
12. In item 10, the above main processing step is A graph neural control differential equation-based spatiotemporal data processing method characterized by further including an initial value generation step for generating initial values for the above-mentioned temporal processing and spatial processing, and training a CDE function including parameters of an initial value generation layer, a node embedding matrix, and an output layer.

Description

Apparatus and Method for Processing Spatial-Personal Data Based on Graph Neural Controlled Differential Equations The present invention relates to a spatiotemporal data processing technology, and more specifically, to STG-NCDE technology that can improve the accuracy of spatiotemporal graph data processing by combining NCDEs for temporal processing and NCDEs for spatial processing into a single framework based on Neural Controlled Differential Equations (NCDEs). Spatio-temporal graph data frequently occurs in real-world applications, ranging from traffic volume to climate forecasting. For example, the traffic volume forecasting task initiated by the California Performance of Transportation (PeMS) is one of the most popular problems in the field of spatio-temporal processing. Time series of the graph Given this, here is a fixed node set, is a fixed edge set, Is It is the point in time when is observed, and is of the node - Time including dimensional input features It is the feature matrix at, and the spacetime prediction is Predicts. For example, given N+1 past traffic volume patterns, predict the traffic volume for each location on the road network during the next S time points. Here, Since is the number of predicted locations and volume is a scalar, i.e., the number of vehicles am. and It does not change over time. That is, the graph topology is fixed, but the node input functions can change over time. Various techniques have been proposed for this task. Meanwhile, NCDE, which is considered a successor to Recurrent Neural Networks (RNNs), is defined by the following mathematical formula. [Mathematical Formula] Here, is a continuous path that takes values in Banach space. The entire trajectory is the path It is controlled over time by the control differential equation (CDE) function for downstream tasks. Tilting is the core of NCDE. CDE theory is stochastic differential equations and Ito calculus It was developed to extend far beyond the Junmartingale settings. In other words, It is reduced to a stochastic differential equation only if the semi-martingale condition is satisfied. For example, the path A general example is the Wiener process in the case of stochastic differential equations. However, in CDEs, the path It does not need to be such a quasi-martingale or martingale process. NODE is a technique that parameterizes these CDEs and learns from data. Furthermore, NODE is equivalent to a continuous RNN and demonstrates state-of-the-art accuracy on many time-series tasks and data. However, methods to combine NCDE technology (i.e., temporal processing) and graph convolution processing technology (i.e., spatial processing) have not yet been studied. FIG. 1 is a diagram illustrating a spatiotemporal data processing system according to the present invention. FIG. 2 is a diagram illustrating the system configuration of a spatiotemporal data processing device according to the present invention. FIG. 3 is a diagram illustrating the functional configuration of a spatiotemporal data processing device according to the present invention. FIG. 4 is a flowchart illustrating a spatiotemporal data processing method based on graph neural control differential equations according to the present invention. Figures 5a and 5b are drawings illustrating the workflow of NCDE according to the prior art and the present invention. FIGS. 6 to 10 are drawings illustrating experimental results related to the present invention. The description of the present invention is merely an example for structural or functional explanation, and therefore the scope of the present invention should not be interpreted as being limited by the examples described in the text. That is, since the examples are subject to various modifications and may take various forms, the scope of the present invention should be understood to include equivalents capable of realizing the technical concept. Furthermore, the objectives or effects presented in the present invention do not imply that a specific example must include all of them or only such effects; therefore, the scope of the present invention should not be understood as being limited by them. Meanwhile, the meaning of the terms described in this application should be understood as follows. Terms such as "first," "second," etc., are intended to distinguish one component from another, and the scope of rights shall not be limited by these terms. For example, the first component may be named the second component, and similarly, the second component may be named the first component. When it is stated that one component is "connected" to another component, it should be understood that it may be directly connected to that other component, or that there may be other components in between. Conversely, when it is stated that one component is "directly connected" to another component, it should be understood that there are no other components in between. Meanwhile, other expressions describing the relationships between