机器学习-白板推导 P5_5
SVD
S = G K G T G T G = I K = [ k 1 0 . . . 0 0 k 2 . . . 0 ⋮ ⋮ ⋱ ⋮ 0 0 . . . k p ] k 1 ≥ k 2 ≥ . . . ≥ k p S=GKG^T \quad G^TG=I \quad K=\begin{bmatrix} k_{1} & 0 &...& 0 \\ 0 & k_{2} &...& 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 &...& k_{p} \\ \end{bmatrix} \quad k_1 \geq k_2 \geq ... \geq k_p S=GKGTGTG=IK=⎣⎢⎢⎢⎡k10⋮00k2⋮0......⋱...00⋮kp⎦⎥⎥⎥⎤k1≥k2≥...≥kp
原始数据为 X X X,中心化矩阵 H = I − 1 N 1 N 1 N T H=I-\frac{1}{N}1_N1_N^T H=I−N11N1NT
对数据进行中心化 H X HX HX
中心化后数据进行 S V D SVD SVD: H X = U Σ V T U T U = I      V T V = V V T = I HX=U\Sigma V^T \quad U^TU=I \;\;V^TV=VV^T=I HX=UΣVTUTU=IVTV=VVT=I
S = X T H X = X T H T H X = V Σ U T ⋅ U Σ V T = V Σ 2 V T S=X^THX=X^TH^THX=V \Sigma U^T \cdot U\Sigma V^T=V \Sigma^2V^T S=XTHX=XTHTHX=VΣUT⋅UΣVT=VΣ2VT
所以:
G
=
V
K
=
Σ
2
G=V \quad K=\Sigma^2
G=VK=Σ2
定义:
T
=
H
X
X
T
H
=
U
Σ
V
T
⋅
V
Σ
U
T
=
U
Σ
2
U
T
T=HXX^TH=U\Sigma V^T \cdot V\Sigma U^T = U \Sigma^2 U^T
T=HXXTH=UΣVT⋅VΣUT=UΣ2UT
T T T和 S S S有相同的特征值
S S S:特征分解,得到方向(主成分),然后 H X ⋅ V → HX \cdot V \rightarrow HX⋅V→坐标
T T T:特征分解,直接得到坐标
T T T主坐标分析(principle coordinate analysis)PCoA
H X ⋅ V = U Σ V T V = U Σ HX \cdot V= U\Sigma V^TV=U\Sigma HX⋅V=UΣVTV=UΣ
U Σ U \Sigma UΣ为坐标矩阵
T U Σ = U Σ 2 U T ⋅ U Σ = U Σ Σ 2 T U \Sigma = U\Sigma^2U^T \cdot U\Sigma=U\Sigma\Sigma^2 TUΣ=UΣ2UT⋅UΣ=UΣΣ2
U Σ U\Sigma UΣ为特征向量组成的矩阵, Σ 2 \Sigma^2 Σ2特征值矩阵