An Approach for Demand Forecasting in Steel Industries Using Ensemble Learning

Table 3

Different machine learning techniques with hyperparameters to be tuned by the grid search algorithm during cross-validation.

Algorithms	Hyperparameters	Explanation	Grid
SVR		Regularization parameter of the error term
	Kernel	Kernel types applied in the algorithm	Linear, polynomial, RBF
	Epsilon	Border of tolerance
	Gamma	Kernel coefficient for rbf
RFR	n_estimators	Number of trees in a forest
	Criterion	Measurement of the quality of a split	mae or mse
	max_depth	Highest depth of the tree
	min_samples_leaf	Least number of instances needed to split an internal node
	min_samples_split	Least number of instances needed to be at a leaf node
MLP	initial_learning_rate	Learning rate value at the starting point of training
	Solver	Used for weight optimization	lbfgs, sgd, Adam
	learning_rate_adjustment	Learning the rate adjustment depending on the cost function’s current value	Constant, adaptive
	hidden_layer_sizes	Layer: Number of layers between input and output layers
		Neurons: Number of hidden layer neurons
	activation_functions	Output of each neuron	Logistic, tanh, relu
	Alpha ( penalty)	Reduces the influence of input parameters
ELM	n_neurons	Number of hidden layer neurons
	activation_functions	Transformation function of hidden layer neurons	relu
	Alpha	Regularization strength
GBR	n_estimators	Number of boosting stages to carry out
	max_features	Number of features while considering the best split	Sqrt
	min_samples_leaf	Least number of instances needed to be at a leaf node
	max_depth	Utmost depth of individual regression estimators
	learning_rate	Shrinks the contribution of each tree
	Loss	Loss function based on order information of input variables	ls, lad, huber, quantile
XGBR	cosample_bytree	Subsample ratio of columns while building each tree
	Subsample	Subsample ratio of training samples
	reg_lamda	regularization On weight
	reg_alpha	regularization On weigh
	min_child_weight	Least sum of sample weight required in a child
	learning_rate	Step size reduction to prevent overfitting
LASSO	Alpha ( penalty)	A constant value that multiplies