Your problem is not very easy to solve, and definitely it does not exist an answer which is for sure correct.

The first thing that come to my mind is that you can probably cluster the data points with a standard clustering algorithm using (number of clusters) = (number of data points you would like to keep). Then you can keep only one point per cluster (probably the centroid of the cluster) and this set will be your new dataset.

Does this make sense?

Cheers

Marco

One option is to use Self Organizing Maps (SOM). SOM is a Neural Network that will help you cluster your data.

M A K

I have similar problem: we are using image features for algorithm selection. The data - features - are very numerous and clumped together. I used various clustering/dimension reducing techniques but unfortunately I was not able to find a general method that is suitable for it all. But you should try to have a look for selections using minimum redundancy and maximum relevance for instance. PCA - also multi-linear PCA can help you to reduce the data size by selecting how much of the information you want to loose. BTW no matter what kind of technique you will apply you will in general loose some data - even by clustering.

I agree with Mr Lukac.

One of the most important stages after feature extraction, is feature reduction.

in this stage you should rank your features to keep important features. This means max dependency and min redundancy between input and target . PCA in Matlab has a criteria that shows after PCA transform how many features you need to keep your precision high.

Also max dependency and min redundancy is usable as fitness function for genetic algorithm that is very good feature selector.

If you want to use PCA, help of Matlab is enough to start. if you want to use GA and mutual information, you can use these papers:

Hanchuan Peng; Fuhui Long; Ding, C.; , "Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy," Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol.27, no.8, pp.1226-1238, Aug. 2005

Ludwig, O.; Nunes, U.; , "Novel Maximum-Margin Training Algorithms for Supervised Neural Networks," Neural Networks, IEEE Transactions on , vol.21, no.6, pp.972-984, June 2010

Ithink you are choose Newral Network Method and Cluster your data, for your problem without doing computationaly intensive operations.

Well, I recommend trying the PCA function in Matlab. In fact there is no unique variable selection technique that's best for all applications and problems.

You can use clustering techniques to reduce your data

a severel method existe and it deponds about the type of your data.

among the well knows clustering method are hirarchical clustering , K means clustering,

try to read a survey and comparative stady about them and choose the most Appropriate to your data.

Thank you all. Based on your experiences there is no variable selection method that works best for all problems. I'll try clustering or t.test. PCA is also interesting since it select variables that are linearly uncorrelated.

Hello Noma,

what is your problem exactly? Do you want to have less samples or less features? Those are two very different things!

Clustering or SOM approaches can be used for reducing the number of samples, while PCA or other dimensionality reduction methods for reducing the numebr of features you have.

Cheers

The problem is topology prediction for membrane proteins. Protein sequences are variable length strings and most classifier can't handle variable length strings. I am using sliding windows to transform these strings into fixed length strings. Therefor for a sequence of length e.g. 100 I generate 100 sub strings of length (windows). let say this is a protein sequence MRHIAHTQ and windows width is 4 then my data set looks like this:

MRHI

RHIA

HIAH

IAHT

AHTQ

AS you can see these samples are very similar and since protein sequences are very long than the size of data set is huge.

PCA is a good and easy choice. but neverthless tht will make u loose some of data.

Some basics might help in selecting data reduction process.

http://www.tdx.cat/bitstream/handle/10803/10479/lozano.pdf;jsessionid=B58F24F9C4015580A825A63631EBA866.tdx2?sequence=1

http://www.stanford.edu/~thairu/07_184.Guest.1sts.pdf

I have a new idea for you to tackle the problem. For data reduction you can reduce the data in 2 ways, feature reduction and data point reduction. I suggest to use Independent Component Analysis- ICA -as a powerful method for feature reduction ( more stronger than PCA). Regarding huge data, I recommend a stochastic method like RANSAC to select a subset of the data.

Maybe you should covert your data to numeric format using your own coding, first.

Try to remove stop words that should help to some extent. You can further use subspace methods (PCA, ICA...)to reduce the dimension.

In data mining, feature selection is a common process. Please specify if your question relates to data set reduction or parameter(attribute) selection. If your question relates to attribute selection, in Rapidminer, multiple steps are available for this task.

I think Subtractive Clustering can be other choice for your work. It usually uses to reduce the number of data in such samples you presented. I had experience using this method for some works and I got good results, Just try it.

It would be easier to put descriptors for each protein or specific sequence. On the other hand, you could identify with the same letter, or put a sequence as derived from the other but explaining all necessary, perhaps Based on graph theory is simpler.

In Bioinformatics, Position Specific Scoring Matrix (PSSM) is used to encode the protein sequence. Since the sequence length greatly varies, moving window approach is suitable. After you have numerically encoded data, you can use any standard data reduction techniques in machine learning as mentioned in previous comments.

Sumit Shrestha are aware of any direct way or tool to convert protein sequences using PSSM?

I believe these links are useful.

http://www.ncbi.nlm.nih.gov/Class/Structure/pssm/pssm_viewer.cgi

http://www.biology.wustl.edu/gcg/psiblast.html

I also found a discussion on generating PSSM in researchgate

https://www.researchgate.net/post/How_to_run_PSIBlast_standalone_and_generate_PSSM_file10

You can use the Singular value decomposition to reduce the dimensionality - it is very popular among text miners. Then, in the reduced space, select representative data by Stratified sampling or some other method.

http://en.wikipedia.org/wiki/Singular_value_decomposition

https://dip.felk.cvut.cz/browse/pdfcache/borovto1_2012dipl.pdf

I start a project as this with my CBR framework in which i implement the similarity measures functions and i remove the similar instances and by a machine learning algorithm i reduce the features. There are many similarity measures function proposed for text data as cosin similarity table, ....etc. Also the traditional approaches of data analyzes (SVA, PCA...etc ) are applied for numeric (real or complex) data, then i propose for you to collaborate with me if you want !!

my email is: [email protected] send to my the data set and i will do my best !!!

I have a big dataset with.Iam looking for a matlab code in order to reduce the instance number  for techniques like  Edited Nearest Neighbor (ENN) ,Repeated ENN ,....

Thank you in advance.

In origin 9 it can be done by importing the data first, then going to analysis, data manipulation, reduce duplicate x data.

Making an statistical sampling of the data set

I am aware that PCA can be used to reduce the length of extracted features, however, I do not know to determine the best length that the PCA would reduce the large data set into. Is there a yardstick for determing the PCA length?