Graph Neural Network Causality Discovery Method
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing causality relationship analysis methods, such as Granger causality analysis, struggle with non-linear variable environments and require manual input of relationships between many input variables, making it difficult to analyze causality in complex systems.
Innovation Solution
A deep neural network training method combining graph neural networks and deep neural networks to automatically extract causality relationships between input variables from training data, using an adjacency matrix to represent causality strengths between variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Granger causality analysis method is used, then causality relationship analysis can be performed, but it cannot handle non-linear variable environments and environments with many variables
Solution Approach 1:
The patent transforms the causality analysis problem from linear parameter space to non-linear parameter space by applying activation functions (ReLU, sigmoid, tanh) in the neural network. This allows the model to capture non-linear relationships between variables while maintaining causality analysis capability through the adjacency matrix representation.
Solution Approach 2:
The patent introduces an adjacency matrix as an intermediary representation that bridges the input variables and output predictions. The adjacency matrix encodes causality relationships between variables, allowing the neural network to learn non-linear patterns while preserving interpretability of causal relationships.
2Measurement precision
If neural network technique is applied to analyze causality relationships, then prediction and recognition performance is improved, but it becomes difficult to interpret inference results
Solution Approach 1:
The adjacency matrix serves as an interpretable intermediary that represents causality relationships between input variables. By training the neural network to learn and output the adjacency matrix, the system maintains both high prediction accuracy and interpretability, as the adjacency matrix provides a clear representation of causal relationships that can be analyzed and understood.
Solution Approach 2:
The patent segments the neural network into distinct functional components: input layer for receiving variables, hidden layers for non-linear transformation, and output layer for producing adjacency matrix and predictions. This segmentation allows the model to maintain interpretability through the structured adjacency matrix output while achieving high predictive performance through deep non-linear transformations.
3Reliability
If graph neural network is applied, then causality relationship analysis is possible, but user has to manually input all relationships between many input variables
Solution Approach 1:
The patent implements self-service by enabling the neural network to automatically learn and discover causality relationships between input variables from training data. The model autonomously constructs the adjacency matrix representation of causal relationships without requiring manual specification, thereby eliminating the burden of manual relationship input while maintaining reliable causality analysis capability.
Solution Approach 2:
The adjacency matrix acts as an intermediary structure that the neural network learns to populate automatically. Instead of requiring manual input of relationships, the network learns to generate the adjacency matrix from training data, bridging the gap between raw input variables and causality relationship representation in an automated manner.
Data Source
AI summary
Provided is a deep neural network training method for detecting causality between input values. The method includes inputting an input value of training data acquired from n input variables to an input layer of a first neural network, which is based on a graph neural network, and calculating a predicted value through an output layer; training the first neural network on the basis of first training information, which is a result of comparing the predicted value to a target value of the training data; receiving an intermediate value in an lth hidden layer (l is a natural number greater than or equal to 1) of the first neural network from a second neural network, which is based on a deep neural network, and calculating an intermediate point value between a point at which the input value is observed and a point at which the target value is observed; and training the first and second neural networks on the basis of second training information based on similarity between the intermediate point value and the input value of the training data.