Machine and Deep Learning Using MATLAB

3 Methods of Improving ML Predictive Models

Accuracy and Robustness of Predictive Models

Multiple classification algorithms are available for training a predictive model on a given dataset. To identify the model that exhibits superior generalizability, a portion of the data can be set aside for validation. Prior to finalizing the model for production, it is worth exploring further refinements. Is it possible to enhance confidence in loss estimates by altering the division between training and validation data? Additionally, can we streamline the model by reducing the number of predictor variables?

Keep in mind that a simpler model is easier to interpret, more computationally efficient, and less prone to overfitting. So far, we have selected a method/classifier to fit the data using a training set, calculated the model loss, validated the model using a test set, and evaluated the validation accuracy, as shown in Figure 3.1.

Figure 3.1 Partition of data into training and testing to tune-up the fitting model accuracy.

Figure 3.2 shows more methods to be adopted in this chapter to better refine the predictability of the classifying model without touching the extremities of the reliability, i.e., neither over-fitting nor oversimplification case. The model ought to be robust and accurate.

Figure 3.2 Methods for further refinement of model predictability in terms of robustness and accuracy. ...