Research On Discriminative Learning Based On Locally Linear Embedding

Feature learning can obtain the feature information of data from high-dimensional data for subsequent data processing,and use less data to display discriminative information in the original data space.Manifold learning is an important method to realize nonlinear feature learning.The classic manifold learning algorithm LLE(Locally Linear Embedding)maintains the local linear relationship of the high-dimensional feature space in the low-dimensional feature space set,and constructs a low-dimensional space mapping to achieve the purpose of feature leaning.However,in the process of feature learning,the LLE algorithm only considers the linear reconstruction error of the data set to maintain the spatial structure of the sample,while ignoring the enhancement of discriminative difference between the features.This thesis proposes two models to make the learned feature set more discriminative by enhancing the difference between features.In the third chapter,an unsupervised discriminative feature learning model called UDLLE(Unsupervised Discriminative Locally Linear Embedding)is proposed,which not only maintains the manifold structure of mapping from high-dimensional space to low-dimensional space,but also the discriminativeness of features is enhanced.Moreover,by constructing discriminative constraints based on the differences between samples,the discriminative information in the feature is strengthen.The optimization of the model feature learning results can be achieved by balancing the proportion of the linear relationship reconstruction error of the samples and the feature discriminative reinforcement constraints in the learning process.In chapter four,a pairwise constraint-guided localized linear embedding discriminative feature learning model called SD-LLE(Semi-Supervised Discriminant Feature Learning Based on Locally Linear Embedding Guided by Pairwise Constraints)is proposed.This model uses the pairwise constraint information that is easy to obtain in practical application scenarios as the supervision information in the model,and guides the direction of feature discrimination enhancement in the learning process,further improving the discriminativeness of the learned features.By adding pairwise constraint supervision information,the similarity between samples of the same cluster and the difference between samples of different clusters are penalized,so that the final learned features have more distinct characteristics between clusters.In the fifth chapter,we design a comparative experiment for the two models and analyze the experimental results.The discrimination of features is evaluated by comparing the feature learning results in common clustering and classification algorithms.And the experimental results in the public high-confidence datasets show that the two models have a great improvement in clustering and classification results in the unsupervised learning and semisupervised learning directions respectively compared with the LLE.