| Machine learning aims to learn models based on training data,and the learned models can be used to predict the outputs for new samples.For classification tasks,traditional supervised learning treats an object in the real world as an "instance & label" pair,where the instance is used to describe the characteristics of the sample,and the label is the category tag of the corresponding object.Traditional single-label learning assumes that each sample is associated with a single and unambiguous category label.However,many machine learning scenarios in real-world applications do not meet the above two assumptions:(1)An instance may be associated with multiple labels simultaneously.(2)The sample in the training set may not have a clear category label description.Based on these two problems,multi-label learning and partial label learning have attracted more and more attention in recent years.The difference between multi-label learning and traditional single-label learning resides in label space.Each instance can be associated with one or more class labels in multi-label learning.In recent years,a large number of multi-label learning methods have been proposed.However,due to the complexity of multi-label learning in label space,existing methods are difficult to perform well on various evaluation metrics.In order to better deal with multi-label classification problems,two new multi-label learning algorithms are proposed in this thesis.Firstly,because of the difference between multi-labeled samples in feature space,we propose a multi-label learning algorithm based on locally adaptive k-nearest neighbor algorithm,which takes into account the differences between samples in different regions of the feature space.In addition,for the difference between the labels of multi-label learning tasks,we propose a multi-label learning method based on label importance analysis.We consider the difference between the predictability and influence of different labels in multi-label learning tasks to determine the label importance.We apply the label importance to two multi-label classification methods,and we achieve better results than the baseline method.Partial label learning aims to induce a multi-class classifier from training examples where each of them is associated with a set of candidate labels,among which only one is the groundtruth label.The common strategy to train predictive model is disambiguation,i.e.differentiating the modeling outputs of individual candidate labels so as to recover ground-truth labeling information.Recently,feature-aware disambiguation was proposed to generate different labeling confidences over candidate label set by utilizing the graph structure of feature space.However,the existence of noise and outliers in training data makes the similarity derived from original features less reliable.To this end,we proposed a novel approach for partial label learning based on adaptive graph guided disambiguation(PL-AGGD).Compared with fixed graph,adaptive graph could be more robust and accurate to reveal the intrinsic manifold structure within the data.Moreover,instead of the two-stage strategy in previous algorithms,our approach performs label disambiguation while training the predictive model simultaneously.Specifically,we present a unified framework that makes ground-truth labeling confidences,similarity graph and model parameters can be jointly optimized to achieve the best results,and we adopt an alternating method to solve this optimization problem.Extensive experiments show that PL-AGGD performs favorably against state-of-the-art partial label learning approaches.This thesis is organized as follows.We review related work about multi-label learning and partial label learning in section 1.Section 2 and section 3 propose two different new multi-label learning methods.Section 4 proposes a novel partial label learning method.And followed by the conclusion in Section 5. |
