Assessing feature selection method performance with class imbalance data

The Fisher Component-based Feature Selection Method

“…A simulation study was conducted to examine the effect of informative and non-informative features on feature selection techniques. Like [4], synthetic data of 30 features with varying imbalance rates and noise (i.e. non-informative features), were generated using the "make classification" library in python's scikit-learn.datasets.…”

“…filter, wrapper, and embedded methods, are widely been used. The filters measure the relevance (e.g., Fisher score, mutual information, correlation [3][4][5]) of a feature with the target. Filters are computationally fast, simple, easy to scale, and generally don't require a learning model [6].…”

“…Greedy (best subset) selection requires 2 , whereas, both forward and backward feature selection techniques require at least models/iterations [2]. Embedded (also known as shrinkage) methods intrinsically maximize the overall performance during the training/learning of a model, like a Least Absolute Shrinkage and Selection Operator (LASSO) or l1 regularization [4]. However, not all the models such as k-Nearest Neighbors (kNN), and decision tree incorporate l1 regularization during their training.…”

“…Authors in [4] ranked the features per the sum of absolute values of the coefficients of the first two principal components. The ranked features were then evaluated with a wrapper.…”

A Comparative Study of Feature Selection Methods for Wind Speed

“…WS/WP is usually predicted based on its historical data and the meteorological data which are commonly represented by extensive time series with multi-dimensions, data noises, and redundancy or lack of attributes. Moreover, using a very complex AI-based forecasting model with a huge volume of input data to train the model is not always feasible because of the limited calculation time and memory [5], [6]. Therefore, approaches for reducing the number of features in the model to eliminate extraneous and redundant data, avoid the loss of relevant information, and reduce computation time are necessary.…”