Skip to main content
. 2021 Jul 31;19:4345–4359. doi: 10.1016/j.csbj.2021.07.021

Table 6.

Major advantages and disadvantages of main learning models described in this work.

Learning Method Pros Cons
Naive Bayes It is simple, does not suffer from outliers, can deal with missing data, is recommended when there are many features and density estimation becomes unfeasible. Assumes conditional independence of input variables given the output label.
Linear Discriminant Analysis Works well on linearly discriminable data. Widely used for Dimensionality Reduction. Assumes that features belonging to the same class have the same Gaussian distribution. Can be sensitive to outliers.
Linear regression Usually the method of choice for small N and/or large M problems. Can be easily regularized with straightforward procedures. Assumes linearity between features and output variables.
Logistic regression and softmax regression Widely used in classification problems for its simplicity as a generalized linear model. Can be easily regularized with straightforward procedures. More robust to outliers compared to LDA. Works well on linear problem. Its extension to non linear problems is computationally expensive and SVM is generally preferred.
Support Vector Machine Works well on linear and nonlinear (Gaussian Kernel function) problems. Not appropriate for large datasets where the number of features exceed the number of observations.
Perceptron learning and ANN Can solve complex problems. Difficult to tune due to the large number of parameters and of possible architectures.
Random Forest Easy to tune. Robust to overfitting. Can provide an estimate of the importance of each feature in the model. Can be slow as computation time increases linearly with the number of trees (inadequate for some real-time applications).
Deep Neural Networks Able to solve complex problems. Difficult to tune due to the large number of parameters/possible architectures.