.png){kind=icon}
.png){kind=icon}
Return
Partial classification and uncertainty estimation under subjective logic
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.113183.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The precise classification algorithms suffer from misclassification when they encounter confusing samples, i.e., out-of-domain (OOD) and imprecise (IM) samples. Consequently, the ability to detect different kinds of confusing samples is a guarantee of models' robustness. To the best of our knowledge, there are two main approaches. Some authors attempt to identify OOD samples through uncertainty estimation. Others assign IM samples to class subsets by performing partial classification. The existing methods focus on one of the approaches and overlook the improvement brought by integrating these two strategies into one method. In addition, in terms of partial classification, some existing methods try to enumerate all the prediction subsets, however, it is impossible when the dataset contains hundreds of classes. In an attempt to alleviate these challenges, based on subjective logic a novelty classification method called PCSL, which can perform partial classification and uncertainty estimation simultaneously is proposed. For uncertainty estimation, a scalar value is quantified to represent the predictive uncertainty. With regard to partial classification, beliefs (i.e., confidence scores) are assigned to potential prediction subsets to aid further decision-making. Experiments demonstrate the excellent performance of the PCSL method as compared to the existing methods. It can be concluded that the PCSL method makes it possible to improve models' robustness by either rejecting confusing samples or assigning confusing samples to prediction subsets.
Keywords:
Partial classification
Uncertainty estimation
Subjective logic
Confusing sample
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W