Improving Robustness of Logistic Regression under Label Noise
Ratnesh Kumar Dubey, Dilip Kumar Choubey · 2021
Administered classification of remotely detected pictures could be a classical strategy to overhaul topographic geospatial databases. The assignment requires preparing information within the shape of picture information with known course names, whose era is time-consuming. To maintain a strategic distance from this issue one can utilize the names from the obsolete database for preparing. As a few of these labels may be off-base due to changes in arrival cover, one should utilize preparing procedures that can adapt with off-base course names within the preparing information. In this paper, the class labels in the training data are commonly assumed to be perfectly accurate. This is not always a justified assumption. Annotators may have made mistakes for example, because the analysis is not obvious or because the annotator is not an expert on the task. It is possible to reduce the impact of labeling mistakes by changing the assumptions the classifier makes. This we explore one way to extend a multiclass logistic regression classifier to account for class-dependent label noise. The vigorous relapse show is assessed on the iris dataset with included manufactured name commotion.