Distance-Based Data Selection to Reduce Annotation Workload
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional annotation methods for supervised machine learning, such as deep learning, impose a heavy burden on users when domain information is not correlated with data or missing, making it difficult to efficiently label large datasets.
Innovation Solution
A data presentation method involving feature extraction, distance calculation, and rearrangement of data in a feature value space, accompanied by dimensionality reduction and clustering, to facilitate efficient labeling by users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If data are rearranged according to domain information (date, time, location, format), then the cognitive burden on user is reduced, but the method cannot be applied when domain information is missing or not correlated with images
Solution Approach 1:
The patent introduces feature values as an intermediary between raw data and domain information. By calculating feature values that represent essential characteristics of data, the system creates a universal mediator that works regardless of whether original domain information is available, thus resolving the contradiction between ease of operation and adaptability
Solution Approach 2:
The patent transforms data from its original form into feature value representations, changing the parameters used for organization. Instead of relying on fixed domain information parameters (date, location), the system dynamically generates feature value parameters that capture data characteristics, enabling the method to work across diverse data types and conditions
2Measurement precision
If many pieces of data are labelled one by one, then accurate labels can be assigned, but the process poses a heavy burden on user
Solution Approach 1:
The patent performs preliminary organization of data by sorting according to feature value distances before the actual labeling process. This preliminary arrangement groups similar data together, allowing users to leverage previous labeling decisions when annotating subsequent data, thereby maintaining high accuracy while significantly improving productivity
Solution Approach 2:
The system provides feedback to users by presenting data in sequences where similar items are grouped together based on feature value distances. This feedback mechanism allows users to infer labels for new data based on previously labeled similar data, reducing the cognitive and time burden while preserving labeling accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In this data presentation method, first, feature values of multiple pieces of data 9 are calculated. Subsequently, a distance d between the pieces of data 9 in a feature value space S defined on the basis of the feature values is calculated. Then, the data 9 is selected from a dataset 90 on the basis of the calculated distance d, and is presented to a user. Thus, the data 9 can be presented to the user in the manner corresponding to the distance d between the pieces of data 9 in the feature value space S. Therefore, the user can efficiently perform a process of labelling each of the multiple pieces of data 9.