Font Size: a A A

Multi-View Multi-Label Learning Via Optimal Classifier Chain

Posted on:2019-12-25Degree:MasterType:Thesis
Country:ChinaCandidate:Y M LiuFull Text:PDF
GTID:2428330545953695Subject:Software engineering
Abstract/Summary:PDF Full Text Request
In the field of machine learning,multi-label learning has evolved from the traditional text classification problem.In multi-label learning,each instance in the training sets is associated with a set of labels,and the purpose of learning is to predict the label sets for the unknown instances by analyzing training instances with known label sets.At present,a series of multi-label classification algorithms have been proposed,which can be roughly divided into two strategies based on their algorithm ideas.One commonly used strategy is the problem transformation method,which converts multi-label classification problems into multiple single classification problems.And then directly apply the traditional single-label classification algorithm to deal with,such as binary classification,multiple classification,and label ranking algorithm.This strategy assumes that each label is independent of each other,and this hypothesis is contrary to the reality and cannot be accurately reflected.The other strategy is adaptive algorithm,which modifies the traditional single-label classification algorithm models so that the models can be applied to the multi-label classification problems.With the deepening of research,researchers have found that combining the correlation among the labels in the design process of the algorithm can improve the classification effect.Both strategies can solve the multi-label classification tasks to a certain extent,but there are still many problems that need to be resolved.With the emergence of multi-view learning as a new form of machine learning,multi-view learning utilizes the complementarity and completeness of multiple views to make learning more efficient.Can multi-label achieve better learning results by utilizing the multi-view learning idea?This paper proposes a Multi-view Multi-label via Optimal Classifier Chain(MVMLOCC),which establishes a multi-label chain classifier for each view of the data sets,then adjusts the weights of the chained learners to predict the unknown instances.When an unknown instance is input,the final set of labels is obtained by multiplying the corresponding weight by multiple classifier chain models.The model makes full use of the correlation be-tween multi-label and the complementarity and completeness of multi-view,and can get better learning performance.Experiments on three challenging real-world multi-label learning datasets,Corel5k,Esp-Game,and Pascal VOC,show that MVMLOCC achieves superior performance to some well-established multi-label learning methods.
Keywords/Search Tags:Multi-Label Learning, Label Correlation, Multi-View Learning, Chain Learning, Weight Adjustment
PDF Full Text Request
Related items