Research On Multilabel Ranking Based On Label Distribution

Multilabel ranking task is the combination of label ranking task and multilabel classifica-tion task.For each instance,multilabel ranking models need to predict the bipartition of the predefined label set into relevant label set and irrelevant label set,and the label ranking over the relevant label set.Multilabel ranking reflects the ambiguity and preference of labels.It is a more specific modeling for real-world applications.However,in existing work,there are few methods designed to handle multilabel ranking task directly.Most of the multilabel ranking algorithms are extended from existing label rank-ing algorithms via calibrated label ranking technique.Therefore,it is difficult to balance the performance of classification and ranking at the same time.In addition,the multilabel rank-ing task has the difficulties of imbalanced classes,limited number of samples,and insufficient annotated information.Therefore it is quite practical to design algorithms to handle multilabel ranking task directly and take into account both classification and ranking performance.Inspired by label distribution model,two multilabel ranking algorithms based on label dis-tributon are proposed in this thesis,multilabel ranking based on label distribution propagation(LDP-MLR)and multilabel ranking based on label distribution PL model(PL-MLR).The for-mer utilizes the similarity information between instances,enhances the original relevant label ranking dataset to label distribution dataset via label distribution propagation process.Then a label distribution model can be learned from the dataset for predicting.The latter combines label distribution with Plackett-Luce probabilistic model,proposes a loss function which can be optimized directly from the original dataset.Then Neural Network is applied as label distri-bution model for training.These two algorithms are applied to two read-world image datasets,and compared with five state-of-the-art mutlilabel ranking methods.Experimental results show that LDP-MLR and PL-MLR perform remarkably better than the compared multilabel ranking algorithms.This thesis is devided into five chapters.The first chapter briefly introduces the research status and deficiencies of multilabel ranking task,and clarifies the motivation and content of this thesis.The second chapter provides the problem definition of multilabel ranking.The related works of multilabel ranking are also reviewed in the second chapter.Several representative al-gorithms are introduced detailly.The third chapter first describes label distribution model,then combines label distribution model with propagation process and PL probabilistic model,pro-poses two multilabel ranking algorithms,LDP-MLR and PL-MLR,respectively.In the fourth chapter,the proposed methods are compared with several state-of-the-art multilabel ranking al-gorithms on two multilabel ranking datasets.Experimental results of ten evalution measures verify the effectiveness of the two proposed methods in terms of both classification and ranking tasks.Finally,conclusions are drawn and the future work are discussed in the fifth chapter.