Composite Machine Learning Model for Unbalanced Data Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine learning models face challenges when dealing with unbalanced datasets, leading to issues of overfitting or underfitting, and existing approaches to address these problems often result in inefficient processing times, high memory usage, and reduced accuracy.

Innovation Solution

A computer-implemented method generates a composite machine learning model by creating multiple training datasets with balanced data subsets from unbalanced datasets, using machine learning algorithms to train separate models for each category, and combining these models for classification and prediction tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional machine learning algorithms are applied to unbalanced datasets, then the model can be trained quickly, but the model suffers from overfitting or underfitting and produces biased outputs

Engineering Contradiction:
Improvemodel accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the unbalanced training dataset into multiple balanced subsets by dividing data across multiple categories. Each subset contains balanced representation of different categories, allowing separate model training for each subset. This segmentation resolves the contradiction by enabling accurate multi-category classification without requiring the entire unbalanced dataset to be processed as a whole, thus maintaining processing efficiency while improving model reliability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by training separate models for each data subset, where each model is optimized for specific category combinations. Instead of training a single model on the entire unbalanced dataset, each local model receives balanced data for its specific subset, improving its accuracy for that local problem while the ensemble of models handles the global unbalanced dataset effectively.

Inventive Principle:
Principle #3Local quality

2Reliability

If model complexity is increased to address underfitting on unbalanced datasets, then training accuracy may improve, but processing time and memory usage increase significantly

Engineering Contradiction:
Improvetraining accuracyVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent segments the training problem into multiple smaller balanced subsets, each trained by a separate model. This segmentation allows each model to be simpler and faster to train, as they only need to handle balanced data for their specific subset rather than the entire complex unbalanced dataset. The time loss is reduced because multiple simple models process data in parallel rather than one complex model processing everything sequentially.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by training models on balanced subsets rather than the complete unbalanced dataset. Each model performs training on a partial balanced subset, which is sufficient for achieving high accuracy on that subset. The excess action is avoided because we only train on the necessary balanced portions rather than attempting to train on all data with increased complexity.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If regularization techniques are applied to prevent overfitting, then model generalization improves, but computational burden on equipment increases

Engineering Contradiction:
Improvemodel generalizationVSAvoidcomputational burden
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent segments the training into multiple balanced subsets, each trained by separate models. This segmentation reduces the computational burden because each model trains on smaller balanced datasets rather than the entire unbalanced dataset requiring heavy regularization. The energy consumption is reduced while maintaining generalization capability through the ensemble of specialized models, each handling a specific subset with appropriate regularization for its size.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4270271A1Method and system for classification and/or prediction on unbalanced datasets
Publication Date: 2023.11.01 SQUIRRO AG
  • EP4270271A1 patent drawingFigure 1
  • EP4270271A1 patent drawingFigure 2A~3
  • EP4270271A1 patent drawingFigure 4~5

AI summary

Disclosed is a computer-implemented method for classification and/or prediction of data samples on unbalanced datasets, the method comprising: receiving unbalanced training dataset; generating at least one first model and at least one second model using unbalanced training dataset, the generation of the at least one first model comprises generating first training dataset comprising first data subset and second data subset of equal amounts of data; the generation of the at least one second model comprises generating second training dataset comprising third data subset and fourth data subset of equal amounts of data, machine learning algorithm(s) are employed for learning from the first and second training datasets; generating composite model using the first model(s) and the second model(s); employing composite model for classification and/or prediction on unbalanced test dataset for generating output, the output includes at least one classified data and/or a prediction for a data sample.