"Support Vector Machine Kernel Weights"
Contributor I
In my analysis I use Support Vecotr Machine (Evolutionary) with Radial Basis Function Kernel. After I finished analysis , I get a result like this:
Total number of Support Vectors: 1124
Bias (offset): -0.011
w[ethni] = -18.791
w[bmi_cat] = -0.728
w[fh_prca] = -79.586
w[pa_cat] = -30.648
w[packyrs_ca] = -108.503
w[ethanol_ca] = 57.925
w[d_lyco_cat] = 24.073
w[p_fat_cat] = 37.213
w[d_calc_cat] = 39.508
w[currsmoke] = -18.024
w[eversmoke] = -8.843
I wonder the meaning of these weights. What they tell me.
Forexample can I interpret the result w[packyrs_ca] = -108.503 as it is the most important attribute (because the weight is highest) that negatively effect the classification result?
Total number of Support Vectors: 1124
Bias (offset): -0.011
w[ethni] = -18.791
w[bmi_cat] = -0.728
w[fh_prca] = -79.586
w[pa_cat] = -30.648
w[packyrs_ca] = -108.503
w[ethanol_ca] = 57.925
w[d_lyco_cat] = 24.073
w[p_fat_cat] = 37.213
w[d_calc_cat] = 39.508
w[currsmoke] = -18.024
w[eversmoke] = -8.843
I wonder the meaning of these weights. What they tell me.
Forexample can I interpret the result w[packyrs_ca] = -108.503 as it is the most important attribute (because the weight is highest) that negatively effect the classification result?
0
Answers
you are right. Attribute weights with a high absolute value play an important role for the classification, but only in that particular model. In order to get a notion about the importance of your attributes it is always necessary to carefully evaluate the model from which you derive your attribute weights. Partitioning attributes in those having high weights and other showing weights near by zero means absolutely nothing when the underlying model has a poor performance.
Cheers,
Helge