Feature Selection Using Semi Discrete Decomposition and Singular Value Decompositions Intisar O. Hussien 1 , Sara Omer 2 1 Arab Open University, Kuwait iothman1@hotmail.com 2 Sudan University of Science and Technology, Sudan saraer.omer@gmail.com Abstract. Nowadays, large amount of digital data is available due to new tech- nologies and different sources of data such as social networks, sensors, etc. There is challenge to deal with this high dimensional data, because query per- formance degrades as dimensionality increases. However, most of this data is redundant. Hence, it can be reduced to smaller number of attributes without significant loss of information. The dimensionality reduction and feature selec- tion techniques can be applied for that. In this paper we compare two tech- niques Semi Discrete Decomposition (SDD) and Singular Value Decomposition (SVD) to select significant features from Hepatitis dataset. We found that SVD is more appropriate than SDD in terms of accuracy and acceptable training time. .Keywords: Dimensionality Reduction, Feature Selection, Semi Discrete De- composition, Singular Value Decomposition. 1 Introduction In recent years data domain contains huge information and large number of features “high dimensions” due to individualized intelligent systems [1]. Examples are multi- media images with thousands of pixels, documents that contain thousands of words, Genomics with thousands of genes. The analysis and query performance degrade rabidly with increase in dimension. This leads to difficulty in interpretation and visu- alization of the data and also computation cost. There is a need to reduce dimensional- ity to be able to apply data mining algorithms on them and achieve fast computation and reduced storage. Dimensionality reduction can be achieved using features extraction by transform- ing from feature space to lower dimension space. This leads to generalization of fea- ture selection so individual features are no longer recognizable [2]. Another possibil- ity is to use features selection approach, which is done by finding a minimum set of features by looking at correlations between the different features and remove the re- dundant ones such that accuracy and performance of the resulting data is as close to the original data as possible.