Term | Definition |
---|---|
Area under the receiver operating characteristic curve (AUC) | A discrimination metric for classification problems, measuring the area under the entire receiver operating characteristic curve. AUC ranges from 0 to 1 with higher values indicating better performance. |
Base-learner | A single, stand-alone statistical or machine learning model built for predicting a continuous or a binary outcome. |
Bootstrapping | Random sampling data with replacement. |
Calibration-in-the-large | A method for measuring the agreement between observed outcomes and predictions for classification problems, where the average predicted probability is compared with the observed event rate. A mismatch indicates that the model over- or underestimates the risk on average. |
Deep neural network | A type of machine learning model that resembles how neurons in human brain work. |
Mean absolute error (MAE) | MAE measures the average magnitude of errors, i.e., the difference between true/observed values and their predictions. Lower MAE indicates better performance. |
Meta-learner | A statistical or machine learning model that uses as input the output of other models (i.e., base-learners), to predict an outcome of interest. |
Multi-layer perceptron (MLP) | The simplest deep neural network model with multiple stacked hidden layers. |
Overfitting | The case when a model fits too closely to the data used to develop the model (training data), but performs badly on new, testing data. |
Permutation feature importance | A method to evaluate the importance of predictors used in machine learning models, by measuring the decrease in model performance when the predictor’s values are randomly shuffled. |
Ridge regression | A statistical regression model which uses a penalized likelihood. The penalty has the effect of shrinking the estimated coefficients so that the model does not yield extreme predictions. |