Research On Problem Transformation Based Partial Label Learning Algorithm

Partial label learning is an important weakly supervised learning framework.In partial label learning,each object is described by a single instance in the input space,associating with a set of candidate labels among which only one is valid.In real-world applications,it is difficult or expensive to obtain an object's ground-truth label,while it is much easier to obtain a set of candidate label for an object.Therefore,partial label learning has attracted the attention of related field.The difficulty of partial label learning lies in the ground-truth label of an object concealed in the candidate labels,and the key to solve the problem is how to disambiguate the candidate labels.From the point of algorithm design,an effective partial label learning algorithm can be designed from two perspectives:algorithm adaptation and problem transformation.In this thesis,we focus on the research of problem transformation based partial label learning algorithm,which mainly include:(1)In order to utilize the potentially useful information of feature space,a feature-aware disambiguation based partial label learning algorithm PL-LEAF is proposed,transforming a partial label learning problem into a multi-output regression problem to solve.The PL-LEAF algorithm first generates the labeling confidence information based on the feature space information of training examples,then utilizes the generated labeling confidence information to learn a regularized multi-output regression prediction model,finally predicts the label which the output value of the regression model is maximum as the label of a test example.(2)In order to decrease the number of candidate labels,ternary error-correcting outputs codes based partial label learning algorithm PL-TECOC is proposed.The algorithm first transforms a partial label learning problem into a series of binary classification problems through a ternary coding matrix,and then ensembles these binary classifiers for prediction.PL-TECOC constructs binary-class training data based only on the labels with non-zero coding,ignoring the labels with zero coding to decrease the number of the candidate labels and thus to reduce the difficulty of partial label learning problem.Experimental results on artificial and real-world datasets show the superiority of PL-LEAF and PL-TECOC,compared with several popular partial label learning algorithms.