Semi-Supervised Classification With Clustering-Based Label Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning methods are ineffective in handling noisy labels, as they assume correct classifications during training, leading to reduced classification performance due to human errors and other factors, and lack mechanisms for noise reduction.
Innovation Solution
A classification system that uses clustering algorithms to define cluster centers from labeled data, computes distances, and adjusts labels based on a ratio value to correct noisy labels, thereby improving classification accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional supervised learning is used to achieve better prediction accuracy, then classification performance is improved, but the cost of labeling training data increases significantly due to required physical experiments, statistical trials, and human labor
Solution Approach 1:
The system performs preliminary clustering on unlabeled data to generate predicted labels before final classification. This preliminary action creates a pre-labeled dataset that can be used for training, reducing the need for time-consuming manual labeling while maintaining classification accuracy through the two-stage process of clustering followed by supervised learning refinement
Solution Approach 2:
The system introduces clustering algorithms as an intermediary step between raw unlabeled data and supervised learning. This intermediary generates initial labels that guide the supervised learning process, reducing the burden of complete manual labeling while preserving the accuracy benefits of supervised approaches
2Reliability
If manual classification is performed to obtain labeled data, then classification models can be trained, but errors due to human fatigue and subjectivity increase label noise
Solution Approach 1:
The system implements feedback mechanisms where clustering results are evaluated and used to identify and correct labeling errors. The algorithm continuously refines labels by comparing clustering assignments with supervised learning predictions, detecting and correcting noise from manual classification through iterative feedback loops
Solution Approach 2:
The system replaces manual mechanical classification with automated clustering algorithms that assign labels based on data patterns rather than human judgment. This substitution eliminates human fatigue and subjectivity while maintaining label availability, using computational methods to determine class assignments objectively
3Productivity
If semi-supervised learning is used to reduce labeling cost, then productivity is improved, but classification performance deteriorates when labels contain noise
Solution Approach 1:
The system performs preliminary clustering to generate initial labels for semi-supervised learning. This preliminary labeling action creates a foundation that guides the semi-supervised algorithm, allowing it to leverage both labeled and unlabeled data effectively while maintaining accuracy even when some labels are noisy
Solution Approach 2:
The system incorporates feedback mechanisms that monitor classification performance and adjust the weighting of labeled versus unlabeled data. When noise is detected in labels, the algorithm reduces their influence and increases reliance on unlabeled data patterns, maintaining productivity while protecting against accuracy degradation
Data Source
AI summary
Data is classified using corrected semi-supervised data. Cluster centers are defined for unclassified observations. A class is determined for each cluster. A distance value is computed between a classified observation and each cluster center. When the class of the classified observation is not the class determined for the cluster center having a minimum distance, a first distance value is selected as the minimum distance, a second distance value is selected as the distance value computed to the cluster center having the class of the classified observation, a ratio value is computed between the second distance value and the first distance value, and the class of the classified observation is changed to the class determined for the cluster center having the minimum distance value when the computed ratio value satisfies a label correction threshold. A classification matrix is defined using corrected observations to determine the class for the unclassified observations.