A meta-analysis on classification model performance in real-world datasets: an exploratory view

Abstract

The No Free Lunch (NFL) Theorem imposes a theoretical restriction on optimization algorithms and their equal average performance on different problems, under some particular assumptions. Nevertheless, when brought into practice, a perceived “ranking” on the performance is usually perceived by engineers developing machine learning applications. Questions that naturally arise are what kinds of biases the real world has and in which ways can we take advantage from them. Using exploratory data analysis (EDA) on classification examples, we gather insight on some traits that set apart algorithms, datasets and evaluation measures and to what extent the NFL theorem, a theoretical result, applies under typical real-world constraints.