多源适应多标签分类框架

发布时间：2018-05-31 04:49

本文选题：特征选择 + 共享特征子空间　；参考：《计算机工程与应用》2017年07期

【摘要】：多源适应学习是一种旨在提升目标学习性能的有效机器学习方法。针对多标签视觉分类问题,基于现有的研究进展,研究提出一种新颖的联合特征选择和共享特征子空间学习的多源适应多标签分类框架,在现有的图Laplacian正则化半监督学习范式中充分考虑目标视觉特征的优化处理,多标签相关信息在共享特征子空间的嵌入,以及多个相关领域的判别信息桥接利用等多个方面,并将其融为一个统一的学习模型,理论证明了其局部最优解只需通过求解一个广义特征分解问题便可分别获得,并给出了算法实现及其收敛性定理。在两个实际的多标签视觉数据分类上分别进行深入实验分析,证实了所提框架的鲁棒有效性和优于现有相关方法的分类性能。
[Abstract]:Multi source adaptive learning is an effective machine learning method to improve the performance of target learning. Aiming at the problem of multi label visual classification, based on the existing research progress, a novel multi-source adaptive multi label classification framework for joint feature selection and shared feature subspace learning is proposed. In the existing graph Laplacian regularization semi supervision The governor learning paradigm takes full consideration of the optimization of the visual features of the target, the embedding of the multi label related information in the shared feature subspace and the use of the discriminant information bridge in many related fields, and it is integrated into a unified learning model. The theory proves that the local optimal solution only needs to solve a generalized feature. The decomposition problem can be obtained separately, and the algorithm implementation and its convergence theorem are given. In the two actual multi label visual data classification, the in-depth experimental analysis is carried out respectively, which confirms the robust effectiveness of the proposed framework and the performance of the classification performance of the existing related methods.
【作者单位】：诺丁汉大学计算机科学与工程学院;浙江大学宁波理工学院计算机与数据工程学院;
【基金】：浙江省自然科学基金(No.LY14F020009) 教育部人文社科基金(No.13YJAZH084) 宁波市自然科学基金(No.2013A610065)
【分类号】：TP391.41;TP181

【相似文献】