Research On Graph-based Semi-supervised Learning And Its Applications

In deep learing,it is very arduous to collect and label samples for training of supervised models.Facing few labeled samples,semi-supervised learning has developed rapidly in recent years.Graph-based semi-supervised learning is one of many semi-supervised learning methods.Compared with other semi-supervised learning methods based on deep learning recently,it has three advantages at least.First,it introduces CNN(Convolutional Neural Network)to assist classification and changes transductive semi-supervised learning to inductive semi-supervised learning,so that makes easy prediction.Second,not only can label propagation help train a more robust network,but this network is beneficial to optimize the graph.Last,graph-based semi-supervised learning uses matrix operations to illustrate the nature of the algorithm,which is highly interpretable.However,graph-based deep semi-supervised learning has many problems,such as excessive reliance on CNN features in construction of graph,the high cost of time and space,lack of applied research currently.Therefore,this paper focuses on these problems and research mainly includes the following three aspects.Aiming at the difficulty of constructing a graph which could reflect the structure of data manifold using single CNN feature,this paper proposes a method of graph-based semi-supervised learning that embeds LBP(Local Binary Pattern)features.In this method,CNN features are extracted by the Res Net-18,and then LBP features are extracted from images.When embedding the LBP features,the idea is to get a graph which is added by the graph constructed by LBP features and the graph constructed by CNN features.But considering the existing conditions,we use an approximate method to achieve the above idea,that is,using the splicing of features to replace the addition of graphs.Finally,the experimental results of this method are obtained on cifar-10 dataset.It shows that when there are fewer labeled samples,the result of embedding LBP features is better,and the accuracy on the dataset has up to 3% improvement,which verifies the method has improved the graph which cannot reflect the structure of the data manifold using CNN features because of the overfitting of the CNN.Focusing on problems of poor discrimination of spliced features and large time consumption of graph construction,this paper proposes a method of graph-based semi-supervised learning based on PCA(Principal Component Analysis).This algorithm uses PCA to find out the subspace of features.After CNN features and LBP features are spliced together,PCA is used to reduce the dimensionality,which can effectively extract the feature principal components and ensure that the graph constructed based on the feature subspace is basically the same as the graph constructed based in original space.Finally,the cifar-10 dataset is used to test this algorithm.The result shows accuracy is greaterly improved with fewer labeled samples.In addition,the k-nearest neighbor searching speed has also been greatly improved after using PCA,which verifies that PCA has improved the quality and efficiency of the graph's construction.For the problem that there are few labeled images in the recognition of newborn pain expressions,this paper applies graph-based semi-supervised learning to neonatal pain expression recognition.In this part,the dataset of newborn pain expression is first established,which mainly includes the collection process of the dataset,the preliminary image processing,and the division of the dataset.Then,under the different labeled samples,the transfer learning strategy is used to obtain results of the graph-based semi-supervised learning in the recognition of neonatal pain expression.The result reflects the effectiveness of graph-based semi-supervised learning method in the neonatal pain expression recognition.