Similarity Graph Modeling for Non-Operational Target Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial intelligence models struggle with predicting when a target system will become non-operational due to the lack of high-quality training data, particularly in situations where tabular data is imbalanced with few instances of the activity to be predicted, making conventional approaches ineffective.
Innovation Solution
A similarity graph-based approach using a graph neural network followed by a gradient boosting machine to leverage information from similar data points, generating a similarity graph and applying it to tabular data to improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional artificial intelligence models are used for prediction, then the system can process data and find patterns, but the prediction accuracy deteriorates due to lack of high-quality training data and imbalanced tabular data
Solution Approach 1:
The patent transforms the prediction problem from a conventional tabular data approach into a graph-structured representation. By constructing a similarity graph where data points become nodes and relationships become edges, the system adds a structural dimension to the data. This graph structure enables the model to capture relationships and patterns that are not apparent in traditional tabular formats, thereby improving prediction accuracy despite limited training data.
Solution Approach 2:
The patent introduces a graph neural network as an intermediary between the input data and the prediction output. This intermediary transforms the raw tabular data into graph representations, processes relationships through the GNN architecture, and then feeds processed features to the gradient boosting machine. This multi-stage intermediary processing allows the system to extract meaningful patterns from imbalanced data more effectively than direct conventional AI approaches.
2Quantity of substance
If manual creation of training data is performed, then the quantity of training data increases, but the time consumption and resource requirements increase significantly
Solution Approach 1:
The system employs automatic similarity computation and graph construction algorithms that self-generate the training data structure without manual intervention. The graph neural network automatically computes similarities between data points, constructs the similarity graph, and processes the training data through learned representations. This self-service approach eliminates the need for manual data creation and programming, significantly reducing time consumption while maintaining data quality.
3Reliability
If graph neural network with similarity graph is used, then prediction accuracy improves by leveraging relationships across data points, but the device complexity increases
Solution Approach 1:
The patent segments the prediction system into distinct modular components: (1) similarity computation module that calculates pairwise similarities, (2) graph construction module that builds the similarity graph structure, (3) graph neural network module that processes graph representations, and (4) gradient boosting machine module that performs final prediction. This segmentation allows each component to be optimized independently and facilitates easier implementation and maintenance despite the overall system complexity.
Data Source
AI summary
Systems and methods for adjusting data processing components. In some aspects, the systems and methods include training a first machine learning model using a similarity graph generated based on training entries to predict whether a target system related to a node in the similarity graph will be non-operational within a future period of time, processing using the trained first machine learning model an updated similarity graph generated based on training and inference entries to predict for each node for the inference entries whether a target system related to the node will be non-operational within the future period of time, processing using a second machine learning model predictions and associated inference entries to predict that a target system related to a node for an entry will be non-operational within the future period of time, and adjusting data processing components related to the target system.