基于三支决策的表情识别集成剪枝算法

A three-way decision-based ensemble pruning algorithm for facial expression recognition

摘要:

集成剪枝通过去除弱学习器和冗余学习器,能显著提升集成系统的表情识别效果。然而,现有方法聚焦于对分类器的接受或拒绝,导致评估信息缺乏可靠性或完整性时,会保留弱分类器或剔除关键分类器。此外,依赖准确率或多样性来评估分类器的优劣,难以全面反映分类器的真实表现。因此,提出基于三支决策的表情识别集成剪枝算法（3WDEP）,引入延迟决策解决分类器评估的不确定性问题。同时,提出“预测偏好”概念,综合预测结果与实际标签的相关性度量、准确率及召回率指标,构建集成剪枝信息系统,全面评估分类器性能。基于熵权法确定指标权重,结合三支决策权衡分类器在不同决策选项下的损失,选出对集成系统贡献最大的分类器进行集成。将召回率作为收益和代价属性,优化集成剪枝效果。实验结果表明,3WDEP能够有效提升表情识别效果,相较于初始集成系统,在FER2013、JAFFE、CK+和KDEF上的准确率分别提高了3.32%、9.39%、1.26%和4.9%。

Abstract:

By removing weak and redundant learners, ensemble pruning can significantly enhance the efficacy of ensemble system-based facial expression recognition. However, existing methods primarily focus on either accepting or rejecting classifiers, which results in the retention of weak classifiers or the exclusion of pivotal ones when evaluation information is unreliable or incomplete. Additionally, relying on accuracy or diversity to evaluate the merits of the classifier is difficult to fully reflect the true performance of the classifier. Consequently, this paper proposed a three-way decision-based ensemble pruning algorithm (3WDEP) for facial expression recognition, which introduced a delayed acceptance strategy to address uncertainties in classifier assessment. Simultaneously, the concept of “predictive preference” was introduced, integrating the correlation measurement between prediction results and actual labels, as well as accuracy and recall metrics, so as to construct an ensemble pruning information system and comprehensively evaluate the classifier performance. The entropy weight method was used to determine the weight of the indicators, and combined with a three-way decision, the loss of classifiers under different decision options was considered to select the classifiers that contributed the most to the ensemble system for integration. Recall was utilized as both a benefit attribute and a cost attribute to optimize the ensemble pruning effect. Experimental results show that 3WDEP effectively improves facial expression recognition performance, and the accuracy improves by 3.32%, 9.39%, 1.26%, and 4.9% compared to the initial ensemble system on FER2013, JAFFE, CK+, and KDEF, respectively.