Authors: Lin Sun, Tengyu Yin, Weiping Ding, Yuhua Qian, Jiucheng Xu

Abstract:

Recently, multilabel classification has generated  considerable research interest. However, the high dimensionality  of multilabel data incurs high costs; moreover, in many real  applications, a number of labels of training samples are randomly  missed. Thus, multilabel classification can have great complexity  and ambiguity, which means some feature selection methods  exhibit poor robustness and yield low prediction accuracy. To  solve these issues, this paper presents a novel feature selection  method based on multilabel fuzzy neighborhood rough sets  (MFNRS) and maximum relevance minimum redundancy  (MRMR) that can be used on multilabel data with missing labels.  First, to handle multilabel data with missing labels, a relation  coefficient of samples, label complement matrix, and label-specific  feature matrix are constructed and implemented in a linear  regression model to recover missing labels. Second, the  margin-based fuzzy neighborhood radius, fuzzy neighborhood  similarity relationship, and fuzzy neighborhood information  granule are developed. The MFNRS model is built based on  multilabel neighborhood rough sets combined with fuzzy  neighborhood rough sets. Based on algebra and information views,  certain fuzzy neighborhood entropy-based uncertainty measures  are proposed for MFNRS. The fuzzy neighborhood mutual  information-based MRMR model with label correlation is  improved to evaluate the performance of candidate features.  Finally, a feature selection algorithm is designed to improve the  performance for multilabel data with missing labels. Experiments  on twenty datasets verify that our method is effective not only for  recovering missing labels but also for selecting significant features  with better classification performance.

Keywords： —Feature selection, fuzzy neighborhood entropy, multilabel fuzzy neighborhood rough sets, MRMR.

Feature selection with missing labels using multilabel fuzzy neighborhood rough sets and maximum relevance minimum redundancy.pdf

查看详情

Thu Dec 23 18:30:00 CST 2021