Contribution to semi-supervised learningapplication to pattern classification

Résumé

The performance of automatic classi cation tasks depends on two main pro- cesses. The rst process is represented by data preprocessing and manifold learning (or dimensionality reduction). The second process is the classi ca- tion task itself. The use of high dimensional data (such as high de nition videos and im- ages) has recently increased, so much that it has become one of the basics of human life today. Our thesis focuses on manifold learning (or dimensional- ity reduction) methods for pattern classi cation. These latter are capable to reduce the size of data while maintaining core content. Since getting a large number of labeled samples is expensive or not accessible in most real appli- cations, our work focuses on semi-supervised methods which use both labeled and unlabeled data samples during the learning phase. In the thesis, we propose several semi-supervised manifold learning meth- ods. Firstly, we propose the label propagation method \Kernel Flexible Man- ifold Embedding" (KFME) which is a kernel version of the classical \Flexible Manifold Embedding" (FME) method. This method overcomes the weakness of FME that appears in case of highly nonlinear distribution of data. Sec- ondly, we propose the \Semi-supervised Flexible Feature Extraction" (SFFE) framework and its kernelized version in order to overcome the limitation of many existing semi-supervised frameworks that are very often transductive, linear, or based on label propagation. SFFE estimates the embedding of sam- ples so that any classi er can use them. Thirdly, we introduce the \Adap- tive Semi-supervised Flexible Feature Extraction" (ASFFE) method that was mainly proposed in order to make the regression function estimated by SFFE more adaptive and elastic. Fourthly, we propose an exponential version of the \Semi-supervised Discriminant Embedding" (SDE) method. This version can overcome the small sample-size (SSS) problem that can disrupt the classical SDE in some cases. Besides the semi-supervised methods, we propose three data-driven strate- gies that can improve the performance of the \Two Phase Test Sample Sparse Representation" (TPTSSR) classi er by automatically estimating its neigh- borhood size parameter which can improve its performance. Experiments performed on di erent benchmark datasets (including facial images, handwritten digits images, objects images and text datasets) prove the performance of our methods in comparison to the state-of-the-art methods.