Tree classifier in singular vertor space

This paper proposes a tree classifier in the local singular vector space of data, named LST algorithm. LST builds oblique decision trees by first transforming local data on the internal nodes to the orthogonal singular vector space and then constructing univariant decision tree nodes in the new space. LST can handle datasets with totally different local and global distribution. Theoretical analysis proves that the time complexity of LST is the same as that of the univariant decision tree algorithms, besides the classification result of LST will not be affected by the arrangement of data samples. Experimental results also show that, compared with the state-of-art univariant decision tree algorithm C4.5 and the well known oblique decision tree algorithms OC1 and CART-LC, LST produces higher classification accuracy, more stable decision tree size, comparable tree construction time as C4.5 and much less than OC1 and CART-LC.