AI Prediction Model Training Data Selection via 2D Coordinate Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data using AI prediction models struggle with selecting data for complex characteristics, leading to overfitting and difficulties in handling multiple attributes, as they often focus on single characteristics and fail to effectively analyze prediction results for complex targets.
Innovation Solution
A method that involves arranging image data with opposing attributes on a two-dimensional plane, using multiple AI prediction models to predict results, converting them into X and Y coordinates, and selecting training data by analyzing distribution using unsupervised learning and local outlier factor analysis to classify data with high and low cohesion, thereby addressing the challenges of complex characteristics and reducing overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If data is selected by applying a threshold to the prediction result, then it is simple to implement for single characteristic, but it is difficult to set thresholds in the case of a complex characteristic
Solution Approach 1:
The patent changes the parameter space by transforming prediction results into a two-dimensional coordinate system where the first coordinate represents one attribute and the second coordinate represents another attribute. This transformation allows complex multi-characteristic data to be visualized and analyzed geometrically, enabling threshold-based selection to work effectively for complex characteristics by defining selection regions in the two-dimensional space rather than using simple scalar thresholds.
Solution Approach 2:
The patent introduces a second dimension by plotting prediction results on a two-dimensional plane with multiple coordinates representing different attributes. This dimensional expansion transforms single-characteristic threshold selection into multi-dimensional region-based selection, allowing simultaneous consideration of multiple characteristics and resolving the limitation of threshold methods when dealing with complex characteristics.
2Quantity of substance
If data with similar characteristics of incorrectly predicted data are included in training, then the amount of training data increases, but it causes overfitting
Solution Approach 1:
The patent applies local quality by analyzing the distribution characteristics of prediction results in different regions of the two-dimensional space. Instead of uniformly including or excluding data, it identifies specific local regions with high data density or abnormal distribution patterns and selectively handles data from these regions. This allows the method to include sufficient training data while avoiding overfitting by excluding data from problematic local regions.
Solution Approach 2:
The patent incorporates feedback mechanisms by analyzing the distribution of prediction results and using this analysis to guide the selection of additional training data. The system continuously monitors prediction performance and data distribution, adjusting the selection criteria based on feedback from the analysis results. This feedback loop ensures that training data is selected optimally to improve model performance without causing overfitting.
3Loss of information
If the prediction result is displayed in the form of a graph assuming it to be time series data, then the shape is visible, but the user may not view two attributes (complex characteristic) together
Solution Approach 1:
The patent transforms the visualization from a one-dimensional time series graph to a two-dimensional coordinate system where each axis represents a different attribute. This dimensional change allows two attributes to be displayed simultaneously and visually correlated, enabling users to view complex characteristics together while maintaining clear visual representation of the data relationships.
4Loss of information
If the prediction result is displayed in the form of a table, then all numerical data is presented, but the user has a lot of numerical data and thus may not visually check the relevance thereof
Solution Approach 1:
The patent transforms tabular numerical data into a two-dimensional visual representation where data points are plotted as coordinates. This transformation converts difficult-to-interpret numerical tables into intuitive visual patterns that can be quickly scanned and analyzed. Users can visually identify relationships, clusters, and anomalies in the data without having to manually examine numerous numerical values, thereby improving visual inspection efficiency while preserving data completeness.
Data Source
AI summary
A method for adding prediction results as training data using an AI prediction model, comprises the steps of: arranging an image having a first opposing attribute, the image being from an actual captured image, on a two-dimensional plane so as to be face-to-face with respect to a specific point on the two-dimensional plane; arranging an image having a second opposing attribute, on the two-dimensional plane so as to be orthogonal, with respect to the specific point, to the image having the first opposing attribute; receiving image data of the first and second attributes respectively arranged on the two-dimensional plane; by means of first and second AI prediction models, predicting respective results for the image data of the first and second attributes; converting the respective prediction results into X, Y coordinates, and cumulating and storing same; and selecting data to be included in training by analyzing the predicted results, and adding the selected data as training data.