Research On New Classes Discovery Learning Based On Semi-supervised Method

The classification faced with open real world application scenarios has received more and more attention in the current research of machine learning,and some such research fields have emerged,such as Zero-Shot Learning,Class-Incremental Learning and so on.This thesis explores a new learning framework faced with open real world,named New Classes Discovery Learning.The New Classes Discovery Learning learns from labeled data belonging to known classes and unlabeled data belonging to known classes and new classes,the learning goal is to construct a multi-class classifier which can classify all classes including all the known classes and the new classes.New Classes Discovery Learning has a great similarity with Semi-Supervised Learning,can be seen as an extension of Semi-Supervised learning to the openness of the real world.However,because all the unlabeled data of semi-supervised learning belong to the class label set of the labeled data,semi-supervised learning methods cannot dig out the new classes knowledge in the New Classes Discovery Learning,nor can learn for obtain a multi-class classifier which can classify all classes including the new classes.In this thesis,a new classes discovery learning method based on the graph-based method in semi-supervised learning is proposed,which improved the famous graph-based method in semi-supervised learning named local and global consistency method.The method enables the original local and global consistency methods have the ability to learn the new classes knowledge of in new classes discovery learning by(1)transferring the distinguished knowledge of the labeled data belongs to the known classes to distinguish all classes,(2)digging out the new classes knowledge with the help of the data distribution of all training data and unsupervised learning skills,(3)establishing the association of all classes with unlabeled data,enables the original local and global consistency methods have the ability to classify new samples with the help of the neighbor samples.The method makes up the shortcomings of the local and global consistency method in solving new classes discovery learning problem,and preserves the advantages of the original method that can learn from the distribution of unlabeled data using the local and global consistency assumption.Experiments have shown that the method made a good use of the unlabeled samples in helping the known classes learning and improving the classification performance and the generalization ability of the model,dug out the new classes knowledge in the unlabeled data and learned from it,and got a satisfactory classification performance on the whole,so it could better solve the new classes discovery learning problem.