An approach to train machine leaning algorithms

Abstract

With the evolution of data mining the trend is to train more complex algorithms. Those sophisticated machine learning algorithms are being used in several fields to take important decisions. Examples in economics include credit scoring models, fraud detection, marketing campaigns, job applications, etc. However, modeling approach choices should not be biased towards complex machine learning algorithms necessarily. Before training machine learning algorithms a working plan should be designed. A strategy could be training a baseline, simple, and interpretable model and, then, rely on more complex ones to ascertain the extent of performance improvements. The purpose of this thesis is to train a logistic regression as a baseline model and challenge it with tree-based ensemble machine learning algorithms to test how much those more complex models improve performance and determine whether it is always worth training complex machine learning algorithms.