Unexpected result when solving a SVM model (Quadratic Programming)

回答済み

Xiangyu Tian

2021年03月07日 01:10

• 
• 
• 

I am trying to use Gurobi to solve the following risk-averse SVM model from a paper:

In the above pic, the model is trying to apply two coherent risk measures on two classes, note that the upper index is not powers, they meant to indicate the first class and second, the lower index means the jth data from that class. So this is a convex quadratic programming, and the only quadratic terms are those in the norm of v squared, here the norm is Euclidean.

And the following is my code:

def c_method(class0, class1, delta, lmd):

    #number of classes

    n = class0.shape[1]

    #number of data in class0

    m0 = class0.shape[0]

    #number of data in class1

    m1 = class1.shape[0]

    m = gp.Model('c_method')

    m.Params.LogToConsole = 0

    v = m.addVars(n,name='v')

    l = m.addVar(name='l')

    z0 = m.addVars(m0,name='z0')

    z1 = m.addVars(m1,name='z1')

    y0 = m.addVars(m0,name='y0')

    y1 = m.addVars(m1,name='y1')

    t = m.addVar(name='t')

    m.setObjective(lmd/m0*y0.sum()+lmd/m0*z0.sum()+(1-lmd)*(t+y1.sum()/(alpha*m1))+delta*np.array(v.select())@np.array(v.select()), GRB.MINIMIZE)

    m.addConstrs((class0.iloc[k,:] @ np.array(v.select()) + z0[k] -l >= 1 for k in range(m0)),name='c0')

    m.addConstrs((class1.iloc[k,:] @ np.array(v.select()) - z1[k] -l <= -1 for k in range(m1)),name='c1')

    m.addConstrs((y0[k] >= z0[k] - 1/m0 * z0.sum() for k in range(m0)),name='c2')

    m.addConstrs((y1[k] >= z1[k] - t for k in range(m1)),name='c3')

    m.addConstrs((z0[k] >= 0 for k in range(m0)),name='c4')

    m.addConstrs((z1[k] >= 0 for k in range(m1)),name='c5')

    m.addConstrs((y0[k] >= 0 for k in range(m0)),name='c6')

    m.addConstrs((y1[k] >= 0 for k in range(m1)),name='c7')

    m.optimize()

    #put all results in a list for later use

    v_=[v[i].x for i in range(n)]

    l_=l.x

    z0_=[z0[i].x for i in range(m0)]

    z1_=[z1[i].x for i in range(m1)]

    sol = [v_, l_, z0_, z1_]

    return sol

 

I was previously using CVXOPT to do this, the results are very close to what is shown in the paper. But as you might know, this package requires matrix inputs, which means the code is not very nice looking, at least to me. So my advisor suggest me to learn about Gurobi, which I personally found very pleasant to code, except that the final result is very strange lol.

Almost all v, which are the coefficients of those features, are extremely small, nearly 0, like 10^(-11). And when I was checking misclassification, one of the classes had no misclassification, and the other one got almost all objects misclassified. But I couldn't find any thing in my code. All data processing codes are identical in two versions (Gurobi one and CVXOPT one), so I assume the function I upload here is what matters only. Could anyone please help me look into it? Thank you so much.

0

• 正式なコメント

  

  Simranjit Kaur

  • Gurobi Staff

  2025年06月23日 02:58 編集済み

  This post is more than three years old. Some information may not be up to date. For current information, please check the Gurobi Documentation or Knowledge Base. If you need more help, please create a new post in the community forum. Or why not try our AI Gurobot?.
• 

  Jaromił Najman

  • Gurobi Staff

  2021年03月08日 13:01 編集済み

  Hi,

  You could use the Model.write() function to write an LP file and check whether the problem looks as you expect. You can write problem files from third party API by using the ResultFile parameter or as described in How do I write an MPS model file from a third-party API?

  Best regards,
  Jaromił

  0
• 

  Xiangyu Tian

  2021年03月08日 22:00

  • 
  • 
  • 

  Hi,

  Thank you for the reply. But I have already tried this, the lp file looks perfect for me, still the answer is somehow very far from perfect. 

  Best,

  Xiangyu

   

   

  0
• 

  Jaromił Najman

  • Gurobi Staff

  2021年03月09日 08:18

  Hi,

  Did you write the LP files from both interfaces, i.e., from Gurobi's Python interface and from the CVXOPT interface? If you write out the two LP files and solve them using the command line tool, do you get the same results when setting ResultFile=solution.sol?

  Best regards,
  Jaromił

  0
• 

  Xiangyu Tian

  2021年03月09日 16:16

  • 
  • 
  • 

  I only inspected LP file written by Gurobi. I don't know if it is possible to write the model with CVXOPT, I have searched for a while but haven't found anyway to do that yet. Do you know how to do this or any links for me to go see for myself? Thank you again.

  0
• 

  Jaromił Najman

  • Gurobi Staff

  2021年03月10日 08:30

  You can write an MPS file as described in How do I write an MPS model file from a third-party API? You can then translate the MPS file to a human readable LP format using the write() function, or via

  gurobi_cl ResultFile=myModel.lp myModel.mps

  The parameter ResultFile makes Gurobi write the problem to an LP file after the optimization problem is done.

  0

投稿コメントは受け付けていません。