| Attribute reduction plays an important role in the development of rough set theory.It aims to eliminate redundant information of data in terms of different requirements.However,in real applications,data usually have the characteristic of multi-granularity.Faced with such challenge,this dissertation will settle the following issues from the viewpoint of multi-granularity.1.In the process of constructing neighborhood,neighborhood rough set only considers the size of radius and the distance among samples.Therefore,when the neighborhood becomes larger,samples with different labels may fall into the same neighborhood.To alleviate the problem,pseudo-label neighborhood rough set is proposed,and then some measures in such model and attribute reductions are re-studied.The experimental results suggest that pseudo-label strategy helps to improve the performances of neighborhood decision system and attribute reductions.2.In the process of heuristic algorithm for computing reducts,different adopted radius will induce different most significant attribute in each iteration of the heuristic algorithm.Hence,the obtained reduct is regarded as unstable.To alleviate the problem,an ensemble heuristic algorithm for computing reducts is proposed.The key step of the algorithm is the voting strategy which is introduced from ensemble learning.The experimental results show that,in terms of the stability,the reducts derived from proposed algorithm are superior to those derived from traditional algorithm in most cases.3.When several Gaussian kernel parameters are required to be considered for computing reducts,the corresponding heuristic algorithm will cost too much time.To reduce the time consumptions,a quick multi-granularity heuristic algorithm is proposed for computing the multi-granularity reducts.Note that only two Gaussian kernel parameters are considered in such algorithm,which are separately derived from the coarsest and the finest granularities.The experimental results suggest that,less time is required to obtain the reducts while the performances of reducts can be guaranteed.4.In multi-label learning,the dimensionalities of the label-specific feature spaces obtained from LIFT are too high to induce classification models quickly.Therefore,attribute reductions in fuzzy rough set model and stable fuzzy rough set model will be performed in those label-specific feature spaces.The experimental results demonstrate that the classification performance of multi-label learning is improved after attribute reductions. |
