Try natural inspired algorithms like ABC

Dear Mitta,

A nice question. optimizing the big features dont satisfy a time cost. I'd prefer to suggest by analyzing the correlation for each features. it can be easier for us to determine in feature selection. Then, if the features are still too many, you can do PCA (mixed the essensial information from each original features into some new features).

Try LASSO. It is by far the most efficient feature selection technique used in Machine Learning.

Dear Mitta,

To make the best choice I suppose you should first know very well your data and the behavior and correlation among variables. Generally, I prefer to use subset selectors algorithms instead of simple variable selection. By doing it you can avoid to keep features with high correlation at the end of the process.

That said, I suggest a greedy approach to search among variables and a subset selector as a evaluation function. Recently I'd used Support Vector Machines as classifier to a Backward Feature Elimination with good results (nevertheless, I was working in a dataset much smaller than yours).

Dear Dhanunjaya Mitta,

You could try a machine learning predictor capable of outputting accuracy and selected features involved in your problem. I'm suggesting the use of Random Forest and its importance (Gini Importance) owing to a better understanding of your problem.

Article Variable selection using Random Forests

I think, since in filter-based feature selection methods, selecting a subset of features is performed based on a specific evaluation criteria without considering the learning algorithms, therefore these kind of methods are proper for processing high-dimensional data. Moreover, it seems in filter-based methods, some methods that are based on the information metrics are better than others. For more details refer to the following paper:

https://dl.acm.org/citation.cfm?id=2307459

Make use of filter approach to reduce some of the dimensionality, then only apply wrapper approach for best feature subset selection. This is so called embedded methods

Hello @all,

The method proposed by my paper is a very fast feature selection method for features over million and large-scale data sets, which can execute in minutes. The time complexity is O(|U|), which ensures a linear increasement of executing time.

Please refer to Article BSSReduce an O(|U|)Incremental Feature Selection Approach fo...

If you want me to help solve your problem, please do not hesitate to contact to me.

Thank you,