Research On Feature Extraction Based On Improved Locality Sensitive Discriminant Analysis

Feature extraction is very important processing step in high-dimension data processingand analyzing tasks, and it is also the most effective and core technology for solving the curseof dimensionality problems which are caused by high-dimension data. In the short passedseveral decades, many researchers have proposed various algorithms for feature extraction.These algorithms can be broadly divided into two classes: linear methods (principalcomponent analysis) and non-linear methods (kernel-based methods and manifold-basedmethods). However, manifold learning is the most popular method for feature extractionwhich has been rapidly developed in the last ten years, and it also has become a hot researchtopic in pattern recognition, machine learning fields and so on.Recently, Locality Sensitive Discriminant Analysis (LSDA) has been proposed as anefficient feature extraction approach. By analyzing the local manifold structure ofhigh-dimensional data, LSDA can obtain a subspace in which the nearby points with the samelabel are close to each other while the nearby points with different labels are far apart.However, because LSDA only takes the local information into consideration, it may fail todeal with the data set which contains some outliers.In order to address this limitation, a new algorithm termed Improved Locality SensitiveDiscriminant Analysis (ILSDA) is proposed in this paper. An intra-class scatter matrix isintroduced into our algorithm, which wills pulls the outlier samples close to their class centerby minimizing the intra-class scatter matrix. Therefore, ILSDA can not only preserve the localdiscriminant neighborhood structure of the data, but also obtains a more compact featureextraction result. Moreover, we also propose an efficient algorithm for performing ILSDA toreduce the computational demand.Finally, the performance of the proposed ILSDA is evaluated on several well known facedatabases and gene expression databases. Extensive experimental results on those datasetsshow the feasibility and effectiveness of our proposed approach.