I'm looking for a strong method to discretization of continuous features. I prefer implemented methods with prepared libraries or functions.
Thanks in advanced
use WEKA its simple and fast. you can use either explorer or Knowledge flow for discretization. yoy can set different parameters according to your need.
KNIME implements the CAIM binner.
http://citeseer.ist.psu.edu/kurgan04caim.html
http://www.knime.org
There is also the LUCS-KDD software for association rule mining
https://cgi.csc.liv.ac.uk/~frans/KDD/Software/
http://citeseer.ist.psu.edu/kurgan04caim.html
https://cgi.csc.liv.ac.uk/~frans/KDD/Software/
http://www.knime.org
I addition to what Oliver listed:
Method 1- very simple method: determine the number of bins (B), divide the range of the data into B to do equal intervals, give value 1 for all numbers located in interval 1, 2 for the second interval and so on.
Method 2- if you know the number of categories: do k-mean clustering on the continuous data to cluster the data into k clusters.
Method 3- if you do not know the number of categories: do Hierarchical clustering on the continuous data to cluster the data into k clusters.
hello friend
There are some methods like
Minimum Splits Based Discretization for Continuous Features
discretizing continuous data include Fayyad & Irani's MDL method, which uses mutual information to recursively define the best bins, CAIM, CACC, Ameva, and many others
Selection of the methods depend on the problem at hand.
All the best
Dr.Indrajit Mandal
use WEKA its simple and fast. you can use either explorer or Knowledge flow for discretization. yoy can set different parameters according to your need.
.
see
M. Boullé. Khiops: a Statistical Discretization Method of Continuous Attributes. Machine Learning, 55(1):53-69, 2004
http://perso.rd.francetelecom.fr/boulle/publications/BoulleML04.pdf
the technique is implemented in the Khiops tool
http://www.khiops.com
.
Could you tell why you want to discretize features? That's a very interesting topic!
I appreciate all comments.
Hello Dear Lendasse,
Many thanks for your comment.
Actually I want to use different machine learning methods and compare them with each other. Some of them like decision tree work with discrete features. So I decided to discretize the features.
How do you measure quality of discretization? Discretization is going to lose some information that is present in continuous values, so, a method where this loss is minimum is a good one. You can apply clustering methods like k-means where sum of squared deviations (loss) is minimized. You can define loss and find the discretization that is going to minimize this loss.
Dear S. Fouladvand
I fully agree with M. Landasse, in that it is always preferrable to ask first what is the actual source of data. But yet there is a pretty robust method for discretization/symbolization that may work in a wide variety of cases: discretization through ordenation. See, for instance, the paper 'Permutation-information-theory approach to unveil delay dynamics from time-series analysis', PHYSICAL REVIEW E, n. 82, 2010.
In Matlab code, for a single 1D signal, it may be as simple as:
s = s(:)';
L = length(s);
n = 5; %I arbitrarily set n=5 to ordene 5 consecutive samples -> segments of s
% are therefore mapped into 5! = 120 symbols
base = n.^(0:n-1)';
for i = 1:L-n+1
[~,ord] = sort(s(j+i:j+i+n-1));
y(i) = (ord-1)*base;
end
To discretize multivaried data, say N simultaneous signals (N channels), a simple adaptation of this method is its application to each channel, thus producing a stream of symbols per channel that can be re-combined to produce a single symbolic stream.
I hope it may help you.
to discretize continous data, just do this:
max_data = max(data)
discr_data = floor(ceiling(data * n_bins) / max_data)
now your data is integers from 0 to n_bins
- Badges
- Science topic