PRELUDE
Regression, a model of learning and finding out some relatively fixed and regular patterns among the chaos. The world of mess and random is somewhat, actually not so eluive as it may seem. There are indeed some underlying but arcane laws or principles or, maybe certain unprovable axioms. If meticulous and scrupulous enough, one would ultimately find the predictability out of the unpredictability.
UNDERSTANDING OF CONCEPTS
Degree of Freedom
二次型矩阵的秩。即为可以自由取值的变量个数。因为n个变量中,又r个线性无关,也就是各自的变动不会影响彼此的值。是最free的。
比如: 看似有4个变量,但是其实r=2, 也就是可以化简成, 就只有两个可以自由取值的。其实就是对角阵 化成
SST的分解
Total sum of squares, 也就是n倍方差,是样本实际值偏离中心的程度。Total, 意味着这个是可分解的量。从残差和误差两个角度去分解。也就是估计值偏离样本实际值的程度+估计值偏离中样本心的程度。
但是SST=SSR+SSE不总是成立,条件:满足OLS的最优情况。因为OLS(least square)时, 最小。对回归参数求一阶导,为0时的参数组,满足SST=SSR+SSE。
R-Square
即:在对于样本中心总的偏离中,有多大程度是由于“回归偏离”引起的。也就是,error 占比越小,regression占比越大,那么样本真实值就越能被regressed出来 模型的拟合度越高。
此时|R|1。