For feature selection, used algorithm is completely dependent on the application and the nature of the processed images. In general, there are many resources for features selection and description, such as:

1. Chapter 4 in Computer Vision: Algorithms and Applications.

2. Chapter 5 in Guide to ?Medical Image Analysis: Methods and Algorithms.

On the other hand. PCA is very good for feature reduction. I think if you use Non-negative Matrix factorization (NMF), it can give you better results (Please, check this paper: The Why and How of Nonnegative Matrix Factorization).

Dear,

i recommand deep learning models

There are multiple approaches for feature selection but one of the best known among them is minimum Redundancy Maximum Relevance (mRMR). You can check following web sites for details:

1. http://shodhganga.inflibnet.ac.in/bitstream/10603/24118/8/08_chapter3.pdf

2. https://medium.com/ai³-theory-practice-business/three-effective-feature-selection-strategies-e1f86f331fb1

3. http://fastml.com/feature-selection-in-practice/

@ ridvan Salih sir will u tell basic difference between wrapper and embedded method

Thank you very much!

Visualize high dimensional featutes using some dimension reduction techniques such as PCA(Principal Component Analysis) OR t-SNE.

Refer this link:

http://www.cs.toronto.edu/~hinton/absps/tsne.pdf

There are lot of methods for feature subset selection. As Mohammed Elmogy correctly pointed out that it completely depends on the application you are looking at. For instance there are Filter methods, Embedded methods and Wrappers methods for supervised learning. For unsupervised and semi-supervised there are quite a few methods available. You can look at the survey paper by: Girish Chandrasheka and Ferat Sahin

Thank Mr Prashant Trivedi , I am interested in unsupervised learning. For more clearification, I have commercial bank's monthly reporting data and want to find an unobservable links between variables(loans, deposits, NPL, staff expences, interest revenue and etc.) or to make clusters of variables wich have influance to each other. I am not sure if it is possible really.

For unsupervised learning there are two different ways that I know of. One is to choose those features which preserves the underlying manifold structure. Other is to using clustering and taking pseudo labels and perform supervised feature selection.

I think you would benefit from looking into Factor Analysis. It is similar to PCA but in my opinion it is superior [1]. It require much more work though [2]. This method played key roles in some of the most impactful discoveries in psychology and other fields, including the discovery of the five factor personality model (FFM).

[1] https://methods.sagepub.com/book/best-practices-in-quantitative-methods/d8.xml

[2] https://aaai.org/ocs/index.php/HCOMP/HCOMP18/paper/view/17914/16925

I my opinion, the best method is the Deep Feature Selection proposed in the article "Deep Feature Selection: Theory and Application

to Identify Enhancers and Promoters"

I have the code in the repository, but it is a bit messy, I apologize, wrote only to verify the method. If will useful you can take advantage of.

https://github.com/mark-rtb/Deep_Feature_Selection/blob/master/keras_app/keras_DeepFeatureSelection.py

also, I often use the method from the library scikit-learn. https://scikit-learn.org/stable/modules/feature_selection

Hello Levani Gafrindashvili , as many of the answers pointed out, there are several methods to perform feature selection in an unsupervised way. I recommend you to look and decide what you want to achieve and identify the structure of your data. There are different methods that can be applied for feature selection on one-class or on a multiclass dataset. There are a lot of books and papers that can help you with that just by searching the term feature selection on google scholar.

On the other hand, to get insights from data, Correlation methods and clustering techniques are a good starting point. I can recommend you WEKA (https://www.cs.waikato.ac.nz/~ml/weka/) to start exploring your options.

Be careful, is not the same to start to discover the relations hidden in your data and then perform feature selection than to prune your data without previous reference and then get the hidden relations.

I recommend you to try Cuttlefish Algorithm

such as in this paper

Conference Paper A New Dimensional Reduction Based on Cuttlefish Algorithm fo...

as an idea, you can use genetic algorithms. because you have several varaibles where too much premitive

as it converges fast so for the calculation time will be fast

and as you do not know the optimal solution because you want to use unsupervised learning

so we adapt your problem to the GA

When talking about dimensionality reduction you have both feature transformation and feature selection.

If I were you I would try different SoA feature transformation methods and on the other hand try to select domenant features by methods as sequential feature selection.