输出结果
设计思路
核心代码
for iStep in range(nSteps):
lam = lam * lamMult
betaIRLS = list(beta)
beta0IRLS = beta0
distIRLS = 100.0
iterIRLS = 0
while distIRLS > 0.01:
iterIRLS += 1
iterInner = 0.0
betaInner = list(betaIRLS)
beta0Inner = beta0IRLS
distInner = 100.0
while distInner > 0.01:
iterInner += 1
if iterInner > 100: break
betaStart = list(betaInner)
for iCol in range(ncol):
sumWxr = 0.0
sumWxx = 0.0
sumWr = 0.0
sumW = 0.0
for iRow in range(nrow):
x = list(xNormalized[iRow])
y = labels[iRow]
p = Pr(beta0IRLS, betaIRLS, x)
if abs(p) < 1e-5:
p = 0.0
w = 1e-5
elif abs(1.0 - p) < 1e-5:
p = 1.0
w = 1e-5
else:
w = p * (1.0 - p)
z = (y - p) / w + beta0IRLS + sum([x[i] * betaIRLS[i] for i in range(ncol)])
r = z - beta0Inner - sum([x[i] * betaInner[i] for i in range(ncol)])
sumWxr += w * x[iCol] * r
sumWxx += w * x[iCol] * x[iCol]
sumWr += w * r
sumW += w
avgWxr = sumWxr / nrow
avgWxx = sumWxx / nrow
beta0Inner = beta0Inner + sumWr / sumW
uncBeta = avgWxr + avgWxx * betaInner[iCol]
betaInner[iCol] = S(uncBeta, lam * alpha) / (avgWxx + lam * (1.0 - alpha))
sumDiff = sum([abs(betaInner[n] - betaStart[n]) for n in range(ncol)])
sumBeta = sum([abs(betaInner[n]) for n in range(ncol)])
distInner = sumDiff/sumBeta
a = sum([abs(betaIRLS[i] - betaInner[i]) for i in range(ncol)])
b = sum([abs(betaIRLS[i]) for i in range(ncol)])
distIRLS = a / (b + 0.0001)
dBeta = [betaInner[i] - betaIRLS[i] for i in range(ncol)]
gradStep = 1.0
temp = [betaIRLS[i] + gradStep * dBeta[i] for i in range(ncol)]
betaIRLS = list(temp)
beta = list(betaIRLS)
beta0 = beta0IRLS
betaMat.append(list(beta))
beta0List.append(beta0)
nzBeta = [index for index in range(ncol) if beta[index] != 0.0]
for q in nzBeta:
if not(q in nzList):
nzList.append(q)
