Attribute Reduction Method Based On Weakly Supervised Learning

In machine learning,supervised learning is a common research method.A prediction model is constructed by learning a large number of training samples,in which each tag output is true.For sample label acquisition,we usually need to spend a lot of cost for manual tagging.Even for some special labels,it is difficult for non-professional personnel to mark such samples,which often consumes a lot of valuable research time of professionals.In addition,irreparable effects are likely to occur as a result of carelessness in the marking process.The emergence of these circumstances makes it difficult to obtain all the true value labels.In order to reduce the occurrence of this kind of situation,we need some methods to deal with the sample set with unlabeled value and with false label value.In recent years,with the introduction of the concept of weakly supervised learning,we have a more intuitive concept.There are three main types of weakly supervision: incomplete supervision,inaccurate supervision and inaccurate supervision.They are usually studied separately,but often appear simultaneously in real life.Based on rough set theory and fuzzy rough set theory,this paper studies the attribute reduction method of weakly supervised learning.1.Weakly supervised learning method is based on rough set.First,the discernibility-matrix based on information table is constructed by similarity relation,and the properties of discernibility-matrix are studied.In real life,due to the cost of labeling data and the negligence of manual labeling,it is difficult to obtain all true labels.In order to solve this problem,the inaccurate and incomplete supervision in weakly supervised learning are studied.Based on the similarity relation,the modified method of weakly supervised learning is given.A new filter MDF,modified decision filter,for dealing with inaccurate and incomplete supervision in weakly supervised learning is proposed by using the approximation operator of rough set under the similarity relation defined.The filter MDF is used to modified the sample,and the attribute reduction method of discernibility-matrix based on the modified decision table is given.2.Weakly supervised learning method is based on fuzzy rough set.First,the discernibility-matrix based on information table is constructed by fuzzy similarity relation,and the properties of discernibility-matrix are studied.Considering that there is such kind of error labeled sample in the data which has no effect on attribute reduction.As the number of samples increases gradually,the sample data with wrong labels will probably become a system independently,Thus,the modified of sample data by the filter is inhibited.Since the fuzzy similarity relation can reflect the similarity between samples more accurately,the upper and lower approximation of the fuzzy rough set model established by the fuzzy similarity relation can suppress the systematic error or noise to some extent.For this reason,this chapter studies the weakly supervised learning based on fuzzy rough set theory,a new filter FMDF is proposed,the filter FMDF is used to modified the sample,and the attribute reduction method of discernibility-matrix based on the fuzzy modified decision table is given.The data sets selected from UCI are used to compare these filters with several common filters.