optimization - Pecuiliar output with Liblinear -
i facing peculiar problem lib-linear package. have 2 levels (+1, -1). have 1 feature takes values $x_1$, $x_2$,..., $x_n$ n points. classifies giving positive weight $w*$ , cost c example. if stack $1$ previous feature make new feature vectors [1 x_i] i=1, 2, ...,n; new problem lib-linear gives following: weight vector [w_1 -w_2]; w_i>0 i.e. weights 1 w_1 , x w_2. cost c1 greater previous cost c.
i understand new feature (1) has no variation throughout , hence weight should automatically go zero.
it minimization problem should give w_1~0 cost c1 @ equal c.
can help?
since have constant input dimension, contribution in decision function constant. liblinear's decision function is
f(x)=sign(w^t*x-rho)
my guess new model corrects term (due non-zero w_1) through rho. can't have idea why w_1 not minimized zero, though. predictions of both models equal?
Comments
Post a Comment