Machine Learning Dataset Optimization via Information Gain

Solution Overview

Problem

Creating large sets of labeled training data for machine learning models is costly and labor-intensive, and existing methods do not optimally allocate resources to identify the most useful data samples for training.

Innovation Solution

A system and method for selecting an optimal representative subset of data instances from a provided dataset, involving an iterative process that identifies and adds data instances with maximal information addition to the subset, until a predetermined threshold is reached, thereby reducing the need for manual annotation and computational resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large quantities of training data are used to create accurate machine learning models, then model accuracy is improved, but the cost and labor required for data annotation increases

Engineering Contradiction:

Improvemodel accuracyVSAvoidannotation cost and labor

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent extracts and selects only the most informative data instances from the full dataset to create a reduced training subset. By identifying and taking out the critical data points that provide maximum information gain, the system achieves accurate model training without requiring annotation of all data, thus reducing labor and cost while maintaining model accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs preliminary analysis of the full dataset to identify and pre-select the most informative data instances before the actual training process. This preliminary action of data selection and prioritization allows the training system to focus only on essential data points, avoiding the costly and time-consuming annotation of redundant data while ensuring adequate training information is available.

Inventive Principle:

Principle #10Preliminary action

2Quantity of substance

If all data instances are annotated manually to ensure comprehensive training data, then data coverage is improved, but the time and resources required increases

Engineering Contradiction:

Improvedata coverageVSAvoidannotation time

Core Design Contradiction:

Quantity of substanceVSLoss of time

Solution Approach 1:

The patent extracts and selects only the most informative data instances from the full dataset to create a reduced training subset. By identifying and taking out the critical data points that provide maximum information gain, the system achieves accurate model training without requiring annotation of all data, thus reducing labor and cost while maintaining model accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The system automatically identifies and selects informative data instances using computational methods, allowing the data to essentially annotate itself through feature extraction and information gain calculation. This self-service approach replaces manual annotation for data selection, dramatically reducing annotation time while maintaining comprehensive coverage of important data patterns.

Inventive Principle:

Principle #25Self-service

3Manufacturing precision

If manual annotation of all data samples is performed to ensure optimal training data selection, then data quality is improved, but the computational and human resources required increases

Engineering Contradiction:

Improvedata qualityVSAvoidannotation system complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent replaces the mechanical manual annotation process with an automated computational system that uses feature extraction and information gain calculations to identify informative data instances. This substitution of mechanical human annotation with automated computational methods maintains data quality selection while significantly reducing the complexity and resource requirements of the annotation system.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system automatically identifies and selects informative data instances using computational methods, allowing the data to essentially annotate itself through feature extraction and information gain calculation. This self-service approach replaces manual annotation for data selection, dramatically reducing annotation time while maintaining comprehensive coverage of important data patterns.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS12299965B2Machine learning training dataset optimization

Publication Date: 2025.05.13 DELL PROD LP

AI summary

A method comprising: receiving a dataset comprising a plurality of data instances; extracting a feature vector representation of each of the data instances in the dataset; choosing a first data instance for adding to a subset of the dataset, wherein the first data instance is removed from the dataset; performing an iterative process comprising: (i) identifying one of the data instances in the dataset which represents a maximal information addition to the subset, based, at least in part, on measuring an information difference parameter between the feature vector representation of the identified data instance and the feature vector representations of all of the data instances in the subset, and (ii) adding the identified data instance to the subset and removing the identified data instance from the dataset, until the information difference parameter is lower than a predetermined threshold; and outputting the subset as a representative subset of the dataset.