Research On Machine Learning Algorithms For Data With Multiple Annotations

In traditional single-label learning,an object is associated with single and explicit labeling information.Under this strong supervision assumption,we can easily obtain classifiers with strong generalization ability.However,objects are often more complicated in real-world scenarios which can not be annotated by a single and explicit label.There are two specific cases which are worth to be investigated:(1)Objects with rich semantics;(2)Objects with ambiguous semantics.In the first case,for objects with rich semantics,valuable information would be lost if each object is only associated with a single label.Therefore,a natural choice is to annotate each object with multiple labels simultaneously to explicitly represent its rich semantics.In the second case,the ground-truth label is difficult to be obtained due to factors such as high labeling cost.Therefore,each object would be annotated with multiple candidate labels among which only one is valid.The above cases correspond to two weakly supervised learning frameworks,namely multi-label learning and partial label learning.It is worth noting that,the multiple labels associated with each object are all valid ones for multi-label learning while are only candidate ones for partial label learning.Under multi-label learning framework,many samples are not only with rich semantics but also with diverse representations simultaneously,i.e.multi-view multi-label learning.For those samples,in this paper,an approach named Simm is proposed with view-specific information extraction.Firstly,Simm jointly minimizes confusion adversarial loss and multi-label loss to utilize shared information from all views.Secondly,Simm enforces an orthogonal constraint w.r.t.the shared subspace to utilize view-specific discriminative information.Extensive experiments on real-world data sets clearly show the favorable performance of Simm against other state-of-the-art multi-view multi-label learning approaches.The ultimate goal of partial label learning is to learn a multi-class classifier from the training examples,while binary decomposition serves as the most straightforward solution for inducing multi-class classification model.Nonetheless,the ground-truth label for each partial label training example is not accessible in the training phase which makes binary decomposition not directly applicable for learning from partial label examples.In this paper,a novel approach named Paloc is proposed which enables binary decomposition for partial label learning by adapting the popular one-vs-one decomposition strategy.Specifically,one binary classifier is derived for each pair of class labels,where partial label training examples with distinct relevancy to the label pair are used to generate the corresponding binary training set.After that,one binary classifier is further derived for each class label by stacking over predictions of existing binary classifiers to improve generalization.Experimental studies on both artificial and real-world partial label data sets clearly validate the effectiveness of Paloc w.r.t.state-of-art partial label learning techniques.This paper concludes 5 chapters.The first chapter introduces the research background,related work and problems to be solved of machine learning algorithms for data with multiple annotations.The second chapter introduces learning frameworks of multi-label learning and partial label learning as well as the related algorithms.The algorithmic details and experiment results of the proposed multi-view multi-label learning algorithm Simm as well as partial label learning algorithm Paloc based on binary decomposition are given in chapter 3 and 4 respectively.Finally,chapter 5 concludes.