Regression model | Full form | Main features |
---|---|---|
MLR | Multiple linear regression | - Explores linear correlations between input variables (Salah et al., 2022) |
Lasso | Least absolute shrinkage and selection operator | - Assigns one of the correlated predictors an elevated weight while minimizing the other correlated predictors to almost zero - Imposes a penalty on the total absolute values of the coefficients, named L1 penalty (Salah et al., 2022) |
Ridge | Regularized inverse depth generating estimators | - Assigns weights that are similar to correlated predictors - Imposes an L2 penalty to the total squared values of the coefficients (Salah et al., 2022) |
Elastic Net | Elastic net regression | - Handles collinear data and prevent overfitting - Combines components of both Lasso (L1) and Ridge (L2) regularization approaches (Malakouti, 2023) |
KNN | K-nearest neighbors | - Makes predictions based on the average of the k-nearest neighbors' majority vote for a given data point - Applies a distance metric (Euclidean distance) which defines "nearest" (Tarek et al., 2023) |
DT | Decision tree | - Identifies the greatest feature and split point at each node using mean squared error (MSE) - Makes judgments by recursively separating the data based on features (Talekar, 2020) |
RF | Random forest | - Combines multiple decision trees to improve prediction accuracy as an ensemble learning method - Entails averaging the predictions made by several trees after they have been trained on arbitrary subsets of the data (Talekar, 2020) |
GBR | Gradient boosting regression | - Reduces the loss function using gradient descent optimization - Utilizes the decision trees which are shallow and have little depth, as weak learners (Tarek et al., 2023) |
AdaBoost | Adaptive boosting | - Creates a series of weak learners, which are usually shallow decision trees, and evaluates their performance using an exponential loss function (Jasman et al., 2022) |
XGBoost | Extreme gradient boosting | - Well-known for its rapidity and effectiveness as a powerful gradient boosting algorithm - Creates a strong regression model by building a sequence of decision trees one after the other, each one fixing the mistakes of the previous one (“POWER | Data Access Viewer”, 2023) |
LightGBM | Light gradient boosting machine | - Well-recognized for its exceptional performance as a gradient boosting framework - Especially effective with the large datasets and provides quicker training times without sacrificing its remarkable accuracy in regression tasks (Malakouti, 2023) |
CatBoost | Categorical boosting | - Combines dynamic learning rates, ordered boosting, and oblivious trees as an advanced gradient boost technique (Jasman et al., 2022) - Gains popularity, especially when working with complex datasets that contain categorical categories |
LSTM | Hyperopt | - Understands relationships in sequential data as a type of recurrent neural network (RNN) - Excels at modeling sequences where long-term context is crucial because of its ability to store and propagate information over extended periods of time (Elsaraiti & Merabet, 2021) |
GRU | Gated recurrent unit | - Develops relationships in sequential data as a type of recurrent neural network (RNN) - Detects long-term dependencies in sequential data while mitigating the problem of vanishing gradients that conventional RNNs encounter (Tao et al., 2022) |