Rectifying Training Data Labels for ML Model Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques for generating labeled training data in machine learning often result in inaccurate labels, leading to poorly performing machine learning models, especially in failure prediction analysis for industrial assets, due to reasons such as lack of correlation in time series data and silent or unrecorded failures.
Innovation Solution
A method is introduced to rectify labels by identifying features relevant to failures using an optimization formulation with rank-one tensor approximation and symmetric cross-entropy loss, and altering uncertain labels based on data similarity measured by a Gaussian kernel function, while maintaining label temporal consistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional techniques are used to generate labeled training data, then the data generation process is simple and fast, but the label accuracy is poor leading to poorly performing machine learning models
Solution Approach 1:
The training data is divided into two distinct regions: a first region containing data points with uncertain or potentially inaccurate labels, and a second region containing data points with reliable labels. This segmentation allows the system to selectively apply complex rectification processes only where needed (in the first region) while maintaining efficiency for the rest of the data.
Solution Approach 2:
An optimization formulation acts as an intermediary mechanism between the two regions. It uses data from the second region (reliable labels) to rectify labels in the first region (uncertain labels) through a mathematical framework that incorporates rank-one tensor approximation and symmetric cross-entropy loss, thereby improving label accuracy without requiring complete re-labeling of all data.
2Reliability
If labels are generated without considering data correlation, then the labeling process is efficient, but the labels fail to capture meaningful patterns leading to poor model performance
Solution Approach 1:
The system performs preliminary analysis to identify and segment data into regions based on label reliability before training the model. By pre-identifying the first region with uncertain labels and the second region with reliable labels, the system can apply targeted rectification strategies that improve model performance without requiring complex post-processing or re-training iterations.
Solution Approach 2:
The optimization formulation changes the parameter space by introducing rank-one tensor approximation and symmetric cross-entropy loss functions. These parameter changes enable the system to effectively measure data similarity and label reliability, transforming the labeling process from a simple annotation task to a sophisticated statistical inference problem that captures meaningful patterns in the data.
3Measurement precision
If all training data is re-labeled to ensure accuracy, then label quality improves, but the time and computational resources required increase significantly
Solution Approach 1:
Instead of re-labeling all training data, the system applies partial action by focusing rectification efforts only on the first region containing data points with uncertain labels. This selective approach achieves sufficient label accuracy for improved model performance while avoiding the excessive time and computational costs associated with complete re-labeling of the entire dataset.
Solution Approach 2:
The system uses data from the second region (with reliable labels) to automatically rectify labels in the first region through the optimization formulation. This self-service mechanism eliminates the need for manual re-labeling of all data, as the system leverages its own reliable data to improve uncertain labels, significantly reducing human time investment while maintaining high label accuracy.
Data Source
AI summary
A method includes obtaining, by a processor set, labeled training data associated with a system; identifying, by the processor set, a first region and a second region in the labeled training data, wherein the first region is associated with a failure of the system and the second region is exclusive of the first region; and creating, by the processor set, re-labeled training data by altering one or more labels of the labeled training data in the first region based on data in the second region.