Labeling Data Selection Using Feature Vector Visualization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for constructing learning data for machine learning models, particularly through crowdsourcing, fail to ensure data quality and often result in biased training due to the indiscriminate labeling of large amounts of raw data, leading to performance degradation.
Innovation Solution
A method involving a service server that derives feature vectors from an original dataset, applies a preset rule to select a subset of these vectors, reduces their dimensions to three or less, and plots them on a visualization interface for user selection, allowing for targeted data extraction and improved data quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If crowdsourcing is used to process large amounts of raw data into learning data, then the time required to construct learning data is reduced, but the quality of the learning data deteriorates
Solution Approach 1:
The patent applies preliminary action by performing feature vector derivation and analysis on raw data before the labeling process. The service server derives feature vectors from raw data, applies selection rules to identify high-quality data candidates, and presents these pre-selected candidates to annotators. This preliminary processing ensures that only data with high potential for quality labeling is processed through crowdsourcing, thereby maintaining data quality while improving efficiency.
2Quantity of substance
If all raw data is used to construct learning data, then the quantity of learning data is increased, but the time and cost for constructing learning data increase
Solution Approach 1:
The patent applies the extraction principle by selectively extracting high-quality data candidates from the entire raw data set through feature vector analysis and selection rules. Instead of processing all raw data, the service server derives feature vectors, applies preset rules to identify the most suitable data candidates, and extracts only these selected candidates for labeling. This extraction approach maintains sufficient data quantity for effective model training while significantly reducing the time and cost associated with processing the complete raw data set.
3Ease of manufacture
If raw data is simply labeled without selection, then the process is simple and fast, but the performance of the trained machine learning model deteriorates
Solution Approach 1:
The patent introduces an intermediary mechanism in the form of a service server that acts as a mediator between raw data and the labeling process. The service server derives feature vectors from raw data, applies selection rules to identify high-quality candidates, and presents these pre-filtered candidates to annotators. This intermediary processing layer maintains the simplicity of the labeling interface for users while ensuring that only high-quality data is labeled, thereby preserving both ease of operation and model performance reliability.
Data Source
AI summary
The present invention relates to a method, a service server, and a computer-readable medium for selectively extracting data for labeling, and more particularly, to a method, a service server, and a computer-readable medium for selectively extracting data for labeling, capable of selectively providing data for effectively training a machine learning model that is desired to be achieved by a user by deriving a plurality of feature vectors for a plurality of data included in an original dataset, applying a preset rule to the feature vectors to select a preset number of feature vectors among the feature vectors, and plotting the selected feature vectors and unselected feature vectors on a plane of three or less dimensions to provide a plotting result to the user.