Semi-Supervised Classification Using Kernel Similarity and Confidence Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised learning algorithms require extensive labeled data for effective prediction, which is costly and time-consuming to obtain, while semi-supervised methods often underutilize unlabeled data for classification tasks.
Innovation Solution
A semi-supervised machine learning approach that computes a similarity matrix using kernel functions to integrate unlabeled and labeled data, iteratively adjusts classifier weights, and employs a sampling technique to enhance the training dataset, allowing for improved classification accuracy with limited labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning algorithms are used to achieve good classification performance, then prediction accuracy is improved, but the cost and time required to obtain labeled data increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-computing a similarity matrix using kernel functions before the iterative classification process. This pre-processing step captures the relationships between all data points (labeled and unlabeled) in advance, enabling the subsequent iterative algorithm to efficiently leverage these pre-established similarities without recalculating them during each iteration, thus reducing the overall time required.
Solution Approach 2:
The patent introduces an intermediary mechanism by using a similarity matrix as a mediator between labeled and unlabeled data. This matrix serves as an intermediate representation that captures the relationships between all data points, allowing the algorithm to indirectly utilize unlabeled data through their similarity relationships with labeled data, rather than directly processing all unlabeled data through expensive labeling processes.
2Measurement precision
If supervised learning algorithms are used to achieve good classification performance, then prediction accuracy is improved, but the cost to obtain labeled data increases significantly
Solution Approach 1:
The similarity matrix acts as an intermediary that enables the algorithm to indirectly utilize unlabeled data through similarity relationships, reducing the need for extensive labeled data while maintaining classification accuracy.
Solution Approach 2:
The algorithm performs self-service by iteratively selecting and labeling its own unlabeled data points based on confidence thresholds and similarity relationships. The system automatically identifies which unlabeled points to label next, eliminating the need for manual intervention to select training samples, thereby reducing both cost and time.
3Quantity of substance
If semi-supervised methods are used to reduce labeled data requirements, then data labeling cost is reduced, but classification accuracy deteriorates due to underutilization of unlabeled data
Solution Approach 1:
The patent applies dynamics by implementing an iterative algorithm that dynamically adjusts the training process. In each iteration, the algorithm selectively adds high-confidence unlabeled points to the training set based on their similarity relationships and predicted labels, progressively refining the classification model. This dynamic adaptation allows the system to effectively utilize unlabeled data without compromising accuracy.
Solution Approach 2:
The algorithm incorporates feedback mechanisms by using confidence thresholds to evaluate predicted labels and selectively incorporating only high-confidence predictions into the training set. This feedback loop ensures that only reliable unlabeled data points are used for training, preventing noise from degrading classification accuracy while still leveraging the abundance of unlabeled data.
Data Source
AI summary
Unclassified observations are classified. Similarity values are computed for each unclassified observation and for each target variable value. A confidence value is computed for each unclassified observation using the similarity values. A high-confidence threshold value and a low-confidence threshold value are computed from the confidence values. For each observation, when the confidence value is greater than the high-confidence threshold value, the observation is added to a training dataset and, when the confidence value is greater than the low-confidence threshold value and less than the high-confidence threshold value, the observation is added to the training dataset based on a comparison between a random value drawn from a uniform distribution and an inclusion percentage value. A classification model is trained with the training dataset and classified observations. The trained classification model is executed with the unclassified observations to determine a label assignment.