Neural network with limited datasets

Hi all,

I am developing back-propagation neural network to classify the incidence of crisis (crisis=1; non-crisis=0) with 15 covariates (a set of macro and economic indicators). I have annual datasets 1970-2012 (42 observations) which I consider it is considerably small for this exercise.

My questions are:

1. Is it okay to proceed for the BP simulation with small datasets and relatively high number of covariates?

2. When I run the simulation, the result keeps changes overtime (In fact, it has similar datasets, training and test data). I just curious why is it happening?

3. Any idea what is the most appropriate classification method to handle small datasets?

Your responses are highly appreciated.

Thanks

Answer 1

[Greg Heath]

 [ I N ] = size(inputs)       % [ 15 42 ]
 [ O N ] = size(targets)      % [  1 42 ]
 Ntrn    = N -2*round(0.15*N) % 30 default (6 val and 6 test)
 Ntrneq  = N*O                % 30 training equations

%For an I-H-O node topology, the number of unknown weights is

 Nw = (I+1)*H+(H+1)*O

 % Therefore, Ntrneq > Nw  <==> H <= Hub where

 Hub = -1+ceil((Ntrneq-O)/(I+O+1))  % 9

Try to minimize H while achieving an adjusted R-squared >= 0.99. I have posted many examples. Search on

 greg patternnet Ntrials R2a

You may also wish to use 10-fold crossvalidation to obtain more precise estimates of error rates.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

The variety of results that you experienced result from the default randomness of trn/val/tst data division and random initial weights. Initializing the RNG to a specified initial state will yield repeatable results.

[rIznaldi]

Thank you Greg for your answer.