Spectral feature selection is a general framework. In this section we show that a number of existing feature selection algorithms are essentially special cases of spectral feature selection. These algorithms include Relief and ReliefF [158], Laplacian Score [74], Fisher Score [45], HSIC [165], and Trace Ratio [130]. These algorithms are designed to achieve different goals. For instance, Fisher Score and ReliefF are designed to optimize sample separability, Laplacian Score is designed to retain sample locality, and HSIC is designed to maximize feature class dependency. We can show that these algorithms actually select feature by evaluating features’ capability to preserve sample similarity in similar ways.