A Study On Fuzzy Rough Decision Trees For Multi-label Classification

Multi-label Classification tasks exist widely in various applications,such as image annotation,text categorization,biology,medical analysis and financial analysis.It's desirable for multi-label classification methods can extract comprehensible and interpretable rules from training samples in biology,medical analysis or financial analysis.However,most of the multi-label algorithms can only learn functions from the training samples.It's worth noting that Multi-label Decision Tree method can extract classification rules.Whereas,Multi-label Decision Tree method assumes labels are independent and overlooks correlations and co-occurrence between labels.In this paper,we propose the following algorithms combining the Fuzzy Rough set theory and considering the Label-Specific Features.Firstly,fuzzy rough decision trees algorithm for multi-label classification,namely ML-FRDT,is put forward.This decision tree adopts the multi-label fuzzy dependency when selecting the attribute of each node.In order to consider the correlation and co-occurrence between labels,this method takes the multi-label problems as multi-class problems and redefines the information gain.The redefined multi-label information gain-ratio is adopted to select the optimal splitting point of each node.ML-FRDT can generate decision trees from symbolic,continuous and fuzzy data.More importantly,intuitionistic and comprehensible rules can be extracted from the multi-label decision trees.Secondly,the multi-label decision tree with label-specific features,named as LIFT-DT,is proposed.In this method,the label-specific features of each label are determined by its information gain-ratio.According to the label-specific features and feature-specific labels,it modifies the multi-label fuzzy dependency in ML-FRDT and adopts it as the evaluation metric when selecting the split point.Meanwhile,LIFT-DT algorithm inherits the advantages from ML-FRDT.It considers the correlation and co-occurrence between labels and can tackle symbolic,continuous and fuzzy data.And it can obtain comprehensible rules from decision trees.In this paper,experiments on several multi-label data sets clearly validate the superiority of ML-FRDT and LIFT-DT against other well-established multi-label learning algorithms.At the same time,the comparison experiments between ML-FRDT and LIFT-DT prove the effectiveness of label-specific features.