Multi-classifier ensemble based on dynamic weights

Abstract

In this study, a novel multi-classifier ensemble method based on dynamic weights is proposed to reduce the interference of unreliable decision information and improve the accuracy of fusion decision. The algorithm defines decision credibility to describe the realtime importance of the classifier to the current target, combines this credibility with the reliability calculated by the classifier on the training data set and dynamically assigns the fusion weight to the classifier. Compared with other methods, the contribution of different classifiers to fusion decision in acquiring weights is fully evaluated in consideration of the capability of the classifier to not only identify different sample regions but also output decision information when identifying specific targets. Experimental results on public face databases show that the proposed method can obtain higher classification accuracy than that of single classifier and some popular fusion algorithms. The feasibility and effectiveness of the proposed method are verified.

SuperCon: Supervised Contrastive Learning for Imbalanced Skin Lesion Classification

“…Lin et al [10] proposes focal loss by reshaping the crossentropy loss in order to reduce the impact of those easily classified samples and majority classes on the loss during the training process. Rather than training a single model, model-level methods normally prepare a pool of models and fuse them to make the final decisions [14,15,16,17]. The contributions of different models to one final decision are normally weighted by each model's reliability, the model confidence on each testing sample or both.…”

“…Finally they fuse the decisions of the filtered classifiers as final decisions. Ren et al [14] determines the classifier reliability by fuzzy set theory, and combines the decision credibility of each testing sample to make the decisions.…”

“…To date, there are a few methods attempting to enhance the performance of CNNs on class imbalanced datasets [5,6,7,8,9,10,11,12,13,14,15,16,17]. In terms of the attention of a deep learning pipeline, the current methods can be roughly divided into three categories: datalevel methods, loss-level methods, and model-level methods.…”

“…For instance, Lin et al [10] proposes focal loss that reshapes the cross-entropy loss so as to reduce the impact of those easily classified samples and majority classes on the loss during the training process. In model-level methods [14,15,16,17], it requires multiple models to be trained on the same dataset and to be fused on the decisions together. The contributions of different models to one final decision are weighted by the model reliability, the model confidence on each sample or both.…”

“…The contributions of different models to one final decision are weighted by the model reliability, the model confidence on each sample or both. For example, given a pool of classifiers, Ren et al [14] determines the classifier reliability by fuzzy set theory, and combines the decision credibility of each test sample to make the final decisions. In summary, most of the current methods aim to tune the training pipeline of CNNs based on the training data class distributions at the data, loss or model levels.…”