Total Pageviews

Wednesday, 9 October 2013

Classification using neural net in r

This is mostly for my students and myself for future reference.

Classification is a supervised task , where we need preclassified data and then on new data , I can predict.
Generally we holdout a % from the data available for testing and we call them training and testing data respectively.  So it's like this , if we know which emails are spam , then only using classification we can predict the emails as spam.

I used the dataset http://archive.ics.uci.edu/ml/datasets/seeds# .  The data set has 7 real valued attributes and 1 for predicting .  http://www.jeffheaton.com/2013/06/basic-classification-in-r-neural-networks-and-support-vector-machines/ has influenced many of the writing , probably I am making it more obvious.

The library to be used is library(nnet) , below are the list of commands for your reference



1.       Read from dataset

seeds<-read.csv('seeds.csv',header=T)

2.       Setting training set index ,  210 is the dataset size, 147 is 70 % of that

   seedstrain<- sample(1:210,147)

3.       Setting test set index

   seedstest <- setdiff(1:210,seedstrain)
 
4.       Normalize the value to be predicted , use that attribute of the dataset , that you want to predict

   ideal <- class.ind(seeds$Class)

5.       Train the model, -8 because you want to leave out the class attribute , the dataset had a total of 8 attributes with the last one as the predicted one

   seedsANN = nnet(seeds[seedstrain,-8], ideal[seedstrain,], size=10, softmax=TRUE)

6.       Predict on training set

   predict(seedsANN, seeds[seedstrain,-8], type="class")

7.       Calculate Classification accuracy


   table(predict(seedsANN, seeds[seedstest,-8], type="class"),seeds[seedstest,]$Class)

Happy Coding !

3 comments:

  1. This is something they were supposed to do long time ago. Listen, there are so many research papers available online for students, so why not taking advantage of those. Not only students can use them by the way. What I mean is that the majority researches have been done already; we just need to work with their results to build a very important data that public can get an access to.

    ReplyDelete
  2. replace "irisdata" with "seeds" and it should work better.

    ReplyDelete

Follow by Email