Comparative study of methods to identify sensitive parameters for improving performance of predictive models

Abstract

Machine learning models map inputs to predictions. Supervised machine learning models learn from a dataset containing several samples or experiments by assigning a weightage to each of the input parameters, commonly referred as features, so as to map to the corresponding target outcome. Different algorithms are used in the learning process, each following a set of rules to achieve the stated objective of mapping features to the corresponding value of target. In this development process, algorithms assign weights to each feature and refine them iteratively to reduce the error between the predicted outcomes with the actual value in the dataset. It is observed that each type of algorithm is based on certain themes such as linear, tree-based, kernel, etc. Each adoption of each of these themed algorithms assigns different weights to features to arrive at the target outcome while reducing the error with the actual value. Iterations alter the weights of parameters until fully tuned and hence there is a need to get reliable weights early in the model development process.