Table 1 Overview of major feature selection algorithm approaches and their characteristics
From: A survey of dimension reduction and classification methods for RNA-Seq data on malaria vector
Feature selection method | Algorithms | Characteristics | Benefits and limitations | Assessments |
|---|---|---|---|---|
Filter Based approaches | Correlation-based feature selection (CBFS) | evaluates a subset by considering the predictive ability of each one of its features individually and also their degree of redundancy (or correlation) | It is feature dependent but slower than univariate techniques | heuristic merit |
Mutual Information | Examined most probable cancer associated genes, to enhance classification accuracy | Evaluates dependencies of features and classes Features contributes redundancy to classification [43] | Symmetric relationship | |
Analysis of Variance (ANOVA) [44] | The dependent variable is continuous and categorized as nominal or ordinal. Its data are normally distributed | It gives overall test of equality of group means It tests against specific hypothesis | Hypothesis test | |
Information Gain [45] | It measures known features of a certain relevant and predicted Information, features that frequently occur in positive samples can be obtained | Its evaluation method based on entropy and it involves lots of mathematical theories and complex theories and formulas about entropy | Ranking | |
Chi-Square [46] | evaluates the correlation between two variables and determines whether they are independent or correlated | |||
Wrapper Based Approaches | Genetic Algorithm [43] | It mimics evolution by taking population of strings to encode possible solutions and combines them to produce more fit | Produces random population search But has lower training time | Crossover and mutation |
Recursive feature elimination method [47] | Backward selection of predictors that fits models and removes weakest features | Has an essential partitioning predictor Ranks features based on the order of their elimination and multicollinearity | Greedy optimization | |
Embedded approaches | Info Gain-SVM [48] | Selects attributes and improves correlation | Reduces the effect of bias resulting from information gain. Adjusts each attribute to allow for the breadth and uniformity of the attribute values | Wavelength |
SVM-RFE [49] | makes implicit orthogonality assumptions, it considers a combination of univariate classifiers The decision function is based only on support vectors that are “borderline” cases as opposed to being based on all examples in an attempt to characterize the “typical” cases | lower risk of overfitting | ranking criterion |