Hi Dhanunjaya,

In my opinion, defining a fixed threshold is difficult (and dependent on the filters actually used).

If you are working on a classification task, one way to determine the optimal number of features (and thus the appropriate threshold) would be to perform the classification. E.g. using cross validation, using a subset of data and see if at some number of best features the classification performance degrades. However, when you do so you should take care that you thoroughly separate your data, i.e. from the aforementioned process you must leave out test data which is independent.

This procedure is related to using a wrapper method. Depending on your classifier, the number of features and the number of realizations maybe you would benefit from using a wrapper method right away instead of the filter. However, the filter approach should be easier to implement, computationally easier and maybe even more robust than a wrapper, which can tend to overtrain.

Greetings, Sebastian

Hi Dhanunjaya ,

I guess the only way to find out is try and error

Thanks

You need to find a baseline that establishes the null hypothesis. You might try a variant of a very old method from radar systems: the constant false-alarm rate (CFAR) filter. It's very straightforward to implement (as it's basically a moving window average).