Open-Set Multiclass Classification Using Similarity Score Ratios
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods struggle to effectively handle open-set scenarios, where test samples may belong to unknown classes not present in the training data, often misclassifying these samples as known classes or failing to reject them properly, especially in high-dimensional spaces.
Innovation Solution
A multiclass classification method that defines a threshold on the ratio of similarity scores rather than absolute similarity scores, allowing for the rejection of unknown samples by establishing a bounded open space of risk, which adapts better to high-dimensional feature spaces without requiring complex parameterization for each dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a threshold is defined on absolute similarity scores to handle open-set classification, then unknown samples can be rejected, but performance degrades in high-dimensional spaces due to difficulty in establishing appropriate thresholds
Solution Approach 1:
The patent transforms the classification threshold from an absolute similarity score to a ratio of similarity scores (between the best-matching class and second-best-matching class). This parameter transformation makes the threshold invariant to the dimensionality of the feature space, eliminating the need for complex parameterization in high-dimensional spaces while maintaining reliable open-set classification.
2Adaptability or versatility
If traditional binary SVM is used for classification, then strong generalization is achieved, but test samples far from training data are incorrectly classified as known classes instead of unknown
Solution Approach 1:
The patent segments the classification problem into two distinct tasks: (1) identifying the best-matching known class among multiple candidates, and (2) determining whether the sample belongs to any known class at all by comparing the ratio of similarity scores to a threshold. This segmentation allows the system to maintain strong generalization for known classes while reliably rejecting unknown samples.
Solution Approach 2:
The patent introduces an intermediary mechanism (the ratio of similarity scores between best and second-best matching classes) that mediates between the strong generalization of SVM and the need for reliable unknown sample rejection. This intermediary ratio provides a nuanced measure that preserves generalization benefits while enabling accurate open-set classification.
3Adaptability or versatility
If multiclass classification is performed using existing methods, then all training classes are covered, but the system cannot properly reject samples from unknown classes
Solution Approach 1:
The patent implements a dynamic decision-making process where the classification outcome is not fixed but depends on the computed ratio of similarity scores. The system dynamically adjusts between classification and rejection decisions based on whether the ratio exceeds the threshold, enabling flexible handling of both known and unknown classes within a unified multiclass framework.
Data Source
AI summary
The proposed method is used for classification in open-set scenarios, wherein often it is not possible to first obtain the training data for all possible classes that may arise during the testing stage. During the test phase, test samples belonging to one of the classes used in the training phase are classified based on a ratio between similarity scores, as known correct class and test samples belonging to any other class are to be rejected and classified as unknown.