Classifier Adaptation via Unlabeled Feature Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning classifiers face challenges in maintaining classification accuracy due to temporal changes in the generation distribution of samples and the emergence of new features not present in the learning data, which are not adequately addressed by existing techniques.
Innovation Solution
A creating device that learns the correlation between features of labeled and unlabeled samples, adding new features from unlabeled data to labeled data, and calculates an importance level to create a classifier that adapts to changing generation distributions and new features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a classifier is created using only labeled learning data, then the classifier can be trained with known features, but it cannot handle new features that appear over time, causing classification accuracy to deteriorate
Solution Approach 1:
The patent applies preliminary action by pre-learning the correlation between labeled and unlabeled data features before new features appear. The system establishes a correlation model in advance that can later be used to infer and incorporate new features from unlabeled data when they emerge, allowing the classifier to adapt without complete retraining.
Solution Approach 2:
The patent uses unlabeled data as an intermediary between the existing labeled data and new features. The correlation learning unit leverages the relationship between labeled and unlabeled data to indirectly capture new features that have not yet appeared in labeled data, serving as a bridge that transfers information about emerging features to the classifier.
2Reliability
If the classifier is retrained frequently to adapt to changing generation distributions, then classification accuracy can be maintained, but the cost and time of relearning increase
Solution Approach 1:
The patent implements feedback by continuously monitoring unlabeled data and using the learned correlation to detect changes in generation distribution. When distribution shifts are detected through the correlation model, the system selectively updates the classifier using relevant information from unlabeled data, rather than performing frequent complete retrainings, thus reducing time loss while maintaining accuracy.
Solution Approach 2:
The patent changes parameters by transforming the approach from frequent full retraining to selective parameter updates. The system modifies the classifier's feature representation parameters by incorporating new features inferred from unlabeled data through the correlation model, allowing adaptive updates with reduced computational cost and time.
3Adaptability or versatility
If new features from unlabeled data are added to labeled data, then the classifier can handle temporal changes, but the complexity of feature correlation learning increases
Solution Approach 1:
The patent applies segmentation by dividing the complex feature correlation learning process into distinct functional units: a correlation learning unit that specifically handles the relationship between labeled and unlabeled features, and a separate classifier creation unit. This modular segmentation reduces overall system complexity by localizing the complex learning task to a dedicated component.
Data Source
AI summary
To create a classifier whose classification accuracy is maintained in consideration of temporal changes in a generation distribution of a sample and a new feature that has not appeared in learning data, a classifier is created in which a feature correlation learning unit learns a correlation between a feature of a sample of labeled learning data, and a feature appearing only in a sample of unlabeled learning data, and a classifier creating unit adds the feature appearing only in the sample of the unlabeled learning data to the feature of the sample of the labeled learning data by using the correlation, and outputs a label associated with an input sample by using the sample of the labeled learning data to which the feature appearing only in the sample of the unlabeled learning data is added.