We are working on high-dimensional features in medical image field. According to our experience, feature selection is a pretty hard step while performing machine learning. Not sure if you are doing high-dimensional features as well. If no, recursive feature elimination (RFE) could be a direct method to improve your model performance through selecting the best combination of features that are related to your outcome. If yes, there would be many methods. We could discuss more.

Yanxi Zhang thank you for your response. We had consider that method for our feature selection. But it is okay for hydrological data?

Z. Shi our data consists of 7 features (year,month,day,time,waterlevel,rainfall_01,rainfall_02)and more than 100,000 rows in .csv file.

Researcher could refer to the paper title: " Feature Selection for Classification: A Review"

In addition, I recommend the Wrapper method using swarm intelligence Metaheuristic -based features selection

Detailed below are several ways which have been explored to do feature selection.

* Filter Methods

Filter methods apply some ranking over features. The ranking denotes how 'useful' each feature is likely to be for classification. Once this ranking has been computed, a feature set composing of the best N features is created. For each of the filter methods described below N was set to 10, 30 and 100. These values were chosen to give an indication of the importance of feature set size.

* Pearson's correlation coefficient.

This method looks at how well correlated two sets of data are. That is, how much does a variation in one data set affect the variation in another. We might want to use this as if there is a high correlation between one feature and the class of the data we are using then this will most likely be good at separating the data and will be useful for classification purposes.

* Mutual Information.

This is an indicator of the shared information between two variables and measures how much knowing one of these variables reduces uncertainty about the other .This works similar to Pearson’s coefficient and is included to see how well it performs in comparison. Again, if there is a lot of shared information between a feature and the class of our labelled data then this is a good indicator that this feature is important and useful in distinguishing members of one class from another.

* Relief.

This method works through a data set looking at the separation capabilities of randomly selected instances.

* Ensemble with data permutation.

The concept of ensemble techniques may be applied to feature selection. In a typical ensemble method, many weak classifiers are created which are combined together to give a result much better than any single classifier could achieve on it's own. This approach may be applied to feature selection by producing many 'weak' rankings which can then be combined together to give a much stronger and more accurate ranking.

* Wrapper Methods

Wrapper methods are so called because they wrap a classifier up in a feature selection algorithm. Typically: a set of features is chosen; the efficacy of this set is determined; some perturbation is made to change the original set and the efficacy of the new set is evaluated. The problem with this approach is that feature space is vast and looking at every possible combination would take a large amount of time and computation. This means that some heuristic search methods must be developed to find optimum sets of features.

* Greedy Forward Search

It works by making changes to the set of features and only keeping the new set if there is an increase in accuracy. Greedy Forward Search works by starting with just one feature and incrementally adding in all the other features. As each feature is added in, the classifier is evaluated with the feature set and the new feature is only kept if there is a notable increase in accuracy

* Exhaustive Search

This means looking at every possible combination of features to find which one gives the best result. It is of course only possible to do this with a small number of features and so some simplification of this problem must be done

the features selection also dependent on the search strategy: some of them are good with small dataset, while the other have the concept of metaheuristic search strategy to escape from local optima as well as overcome the limitation of greedy selection strategy.