Training Data Selection Using Bipartite Graph Attribute Balancing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for selecting training data for deep learning models suffer from attribute-specific bias and imbalance, leading to inaccurate training and performance issues.
Innovation Solution
A method and device that utilize a bipartite graph to select training data based on individual attributes, ensuring each attribute type is within a predetermined deviation threshold, using optimization algorithms like linear and integer programming to achieve balanced data selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If random sampling method is used to select training data, then the selection process is simple and fast, but attribute-specific bias and imbalance occur in the selected data
Solution Approach 1:
The patent transforms the data selection problem from random sampling to optimization-based selection by changing the selection criteria parameters. It introduces cardinal number constraints for each attribute type and uses deviation thresholds to control the balance between different attribute categories, thereby eliminating attribute-specific bias while maintaining selection efficiency through automated optimization.
Solution Approach 2:
The patent replaces the mechanical random sampling process with an optimization algorithm system. By substituting the random selection mechanism with mathematical optimization (linear programming or integer programming), the system achieves both efficiency and reliability by automatically balancing attribute distribution according to predefined constraints.
2Reliability
If vector quantization method is used to select training data, then some attribute-specific bias is alleviated, but data bias and imbalance problems still occur
Solution Approach 1:
The patent segments the training data selection process by attribute types, creating separate cardinal number constraints for each attribute category. This segmentation allows independent control over the distribution of each attribute type while maintaining overall data balance, addressing the limitations of vector quantization's holistic approach.
Solution Approach 2:
The patent performs preliminary action by pre-defining the target cardinal numbers and deviation thresholds for each attribute type before the optimization process. This preliminary configuration enables the optimization algorithm to directly compute the balanced selection without iterative adjustments, reducing complexity while maintaining reliability.
3Reliability
If optimization algorithm is used to select balanced training data, then attribute-specific bias is reduced, but computational complexity increases
Solution Approach 1:
The patent changes the computational approach by formulating the data selection as a parameter-constrained optimization problem. By defining clear parameters (cardinal numbers, deviation thresholds, attribute types) and using standard optimization algorithms, it achieves reliable balanced selection with manageable computational complexity through structured problem formulation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for selecting training data for training a deep learning model is provided. The method includes steps of: (a) obtaining one or more individual attributes each of which corresponds to each of a plurality of training data included in total training data, and generating a bipartite graph by matching each of the plurality of training data included in the total training data with the individual attributes; and (b) selecting n training data among the total training data, by referring to the bipartite graph, wherein the n is a target number of the training data to be used for training the deep learning model, and wherein the training data selecting device selects the n training data to be used for training the deep learning model such that each cardinal number of each of the individual attributes matched with the n training data is within a predetermined deviation threshold.