爱丁堡大学留学生指导assignment-Credit Scoring and Data Mining

University Of Edinburgh
Credit Scoring and Data Mining

Workshop 3: Other techniques for building scorecards

Write out the  linear programme that builds a scorecard on the following data. We are interested in the scorecards that
a) minimise the sum of the absolute errors
b) minimise the maximum absolute error
c) minimises the number of misclassified applicants

 
Applicant no Under 25 Owner With bank for>2 years Good/Bad status
1 Y N Y G
2 N Y Y G
3 Y N N B
4 N Y N G
5 N N N B
.
If you wanted to ensure that under 25s got a higher score than older applicants how would you ensure that?
If you felt that time with bank was more important than either of the other two characteristics, how would you ensure that?

2. In the neural network approach, assume the network has two input nodes ( under 25 and owner), two nodes in the hidden layer and one output node ( good/bad). One wants to minimise the average value of the square of the errors on the output nodes over all training cases.
Assume at each node j the output function is
yj =1/(1+e-u )   if input is u =  y kwkj  where sum is over all nodes k inputting to j.

Use the data in question 1 and starting with all weights wjk  = 1 find what are the values yk(t) at each node for the first two cases (t=1,2), and compare the actual outcome  at then output node o¬j (t) with the predicted value yj(t) after training case t has been entered.

Weights going into output node are then adjusted by  dk(t)y k(t)  where
k(t,u) = (actual(ok (t))-predicted(yk (t))) (e-uk(t) /(1+e-uk(t) ) 2 )  and d is some training rate (make it 0.5)

For the inner nodes at level c, first the errors are calculated from those at level c+1 by
 k c(t) = (e-u(t) /(1+e-u(t) ) 2 )  where u(t) =  wkj  j c+1(t),
 The weights wkj connecting a node in the c and c+1 later are adjusted by djc+1(t)y kc(t) 

 Do these calculations for the first two applicants in question 1 to see what happens to the weights