Neural Networks

vaishaks89

Dear All,
I am using neural networks in rapid miner. What kind of data can be used with neural networks?
I have a data in the form of a table shown below
Admin-region  Admin-Team  Invoice-value  Sector            Organization             Paid-band
Region 1           Team 1            30$               Medical            Sam medicals             Within 30-50 days
Region 2            Team 2            500$            Healthcare        NHS                            Within 0-30 days
A company which performs service to its customers will be generating an invoice. This invoice will normally be paid by customers in say
0-30 days 30-50 days 50-80 days and so on.
1) Is this type of data suitable for neural network? I  doubt because as far as I know neural nets use neumeric data.

2) If we have 10 customers it must be able to  tell that 4 customers will be paying lately
customer 1 will take 0-30 days
customer 2 will take 50-60 days and so on.

MariusHelf

Hi,

you can use only numerical data with the Neural Net. However, you can convert nominal data to numerical data e.g. with the Nominal to Numeric operator (e.g. dummy coding). Furthermore, you could use different learning schemes for classifcation, neural nets are only one option out of dozens.

If you have multiple labels, you could use e.g. Polynominal by Binominal classifcation to use the Neural Net for polynominal classifictaion.

By the way, did you check out our tutorial page? -> http://rapid-i.com/content/view/189/212/lang,en/

Best,
Marius

vaishaks89

THank you very much for the prompt response.

I did try using nominal to neumeric operator. I found that it assigns numbers to words. How much can this conversion be relied when it comes to developing an accurate prediction model?
What I did was to convert Nomainal to neumeric and then use an algorithm called GHA which converts these numeric values into floating point or decimal values
These were then input to the neural nets.
Can such a conversion be relied? Or is it correct? The model gives me an accuracy of 74% with training data. I am yet to use the testing data for the algorithm

MariusHelf

Hi,

you should change the coding type of Nominal to Numeric to "Dummy Coding". Leaving it at "Unique Integers" does not make sense in 95% of all cases. When using dummy coding, you can trust the resulting dataset.

Whether a performance of 74% is good or bad depends heavily on the data and on the use case.

Best,
Marius

vaishaks89

Dear Marius,
I would like to come  back to the question again.
Right now I have a table with columns like these given below. I need to use Neural Nets for this problem. And these columns are its inputs.

Gender    Age      Maritial Status  No of delayed payments  No of policies  No of children
Male      25-35      Married                                                        1                      >1C
Female  <45            Single                    >3                                    >0                  0
Female    25-35      Divorced                                                  >0                  >1C
Male        35-45      Single                    >3                                  >0                    0

By converting these nomainal values to neumeric I have another table which has values like this
0  0    0    0      0  0
1  1    1    1    1    1
1    0    2    0    1  0
0  2      1    1    1  1
With this kind of data neural nets gives me an accuracy of 83%. But how much of it can be relied. I mean I am confused at the logic. When some other data is given the same configuration gives very low accuracy. Is there any explanation with regard to this? 
Is such a data suitable for  neural networks?
Any help would be appreciated.
Thanks

MariusHelf

How did you evaluate your model? Did you a cross-validation, or did you test in on the training data? Generally speaking, your data should be suitable for neural nets, but you should as well try other algorithms such as the SVM and compare the performance. Most important for this: validate your models correctly with a cross-validation (operator X-Validation in RapidMiner).

vaishaks89

Well,
  I evaluated the model using split validation technique. I haven't considered using cross validation so far. Thank you very much for letting me know and for patiently replying my queries.
So given the data as above I am asked to predict whether the same set of people would be active or not the next year. So basically I need the network to predict  the number of active and non active customers.

You have told me that this data is suitable for neural networks.
My question is: Whether if this is the right way to convert the data before giving it to neural network.
Normally in video tutorials most poeple have used neumeric data. But in my case, I am using data like GEnder, Age, Maritial status etc. I am converting them using some operator. Is that the proper way to convert such a data? Or is there any other method to give such inputs to neural networks

PS: I will try using SVM as you already suggested and let you know the results
Regards
Vaishak 

 

jokosdr

Dear All,
I have inflation data with 11 attributes, 10 attributes as attributes as input and 1 output. I use a Hybrid PSO-NN, but there is a problem that is not optimal results. Is it true that the PSO-NN results are more accurate than the NN alone? I either make or model of the process? Please help.