Research On Active Learning Method Based On Fuzzy Set Theory And Its Application

Active learning is an algorithm to reduce the huge cost of marking a large number of unlabeled data.Traditional active learning is often aimed at clear single-label data,and only considers the uncertainty of the samples on the classifier when making sample selection.However,the real data is often fuzzy in some feature attributes,and some data also have multi-label.In this context,this thesis proposes an improved active learning framework based on fuzzy sets and multi-label learning theory,and applies the algorithm to the marking of coal mine safety situation data.The main research contents are as follows:1.A single tag active learning method based on fuzzy set theory is proposed(FTAL)Firstly,this thesis construct information measures that combine uncertainty,inconsistency,representativeness,and diversity,which aimed at the ambiguity involved in the active learning framework.Secondly,this thesis incorporates fuzzy information entropy into the analytic hierarchy process to assign reasonable weights to multi-criteria information measures.Finally,the fuzzy comprehensive evaluation model is used to select the samples with excellent evaluation results for expert marking.Comparative experiments on UCI public data sets show that the FTAL proposed by the ontology on indicators such as Accuracy and F1 value is better than several other active learning algorithms in comparison algorithms.2.A multi-label active learning method based on fuzzy set theory is proposed(ML-FTAL)Firstly,the thesis modified the uncertainty and inconsistency criteria for single-label data in FTAL to adapt to multi-label data.Secondly,this thesis considers the correlation between the sample and the label,as well as correlation between the labels in the label space set.Finally,The experimental results on Mulan open multi-label data set show that the ML-FTAL proposed in this thesis is superior to other comparison algorithms in Hamming Loss and One-error indexes,which can not only maintain the effectiveness of sample selection,but also ensure the marking efficiency of multi-label data.3.Research on application of active learning in coal mine safety situation data markingThis thesis firstly conducts background research on coal mine safety,and then analyzes the characteristics of coal mine safety situation data.After preprocessing some data,this thesis applies the active learning method to mark the coal mine safety situation data.Finally,the experiment results show that the active learning method proposed in this thesis has achieved better results.