scikit-learn 或 statsmodels 中线性回归的调整参数的限制范围

Question

Is it possible to limit the bounds of tuning parameters for linear regression in scikit-learn or statsmodels, eg in statsmodels.regression.linear_model.OLS or sklearn.linear_model.LinearRegression?

http://statsmodels.sourceforge.net/devel/generated/statsmodels.regression.linear_model.OLS.html

http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LinearRegression.html

EDIT :

---

scipy 0.17 includes scipy.optimize.leastsq with bound constraints:

http://docs.scipy.org/doc/scipy-0.17.0/reference/generated/scipy.optimize.least_squares.html#scipy.optimize.least_squares

Ideally what I'm looking is to minimize objective error function and also minimize the change of tuning multiplier parameters from default value of 1.0. This is likely part of objective function.

Note that this is the list of options that worked for my box bounds:

method='trf' or 'dogbox'
loss='cauchy'
f_scale=1e-5 to 1e-2

Answer 1

Not sure what you mean by "limit the bounds of tuning parameters".

• If you'd like the result components to lie in pre-specified ranges, you could try scipy.optimize.least_squares , which solves

  minimize F(x) = 0.5 * sum(rho(f_i(x)**2), i = 0, ..., m - 1) subject to lb <= x <= ub

• If you're concerned about the result components being too large in magnitude due to colinearity, you can try sklearn.linear_model.Ridge (or one of the other regularized linear regressors there).