Continuous Optimizers for Automatic Design and Evaluation of Classification Pipelines

Nowadays, a big pool of different machine learning components (i.e., algorithms and tools) exists that are capable of predicting various decisions in different problem domains successfully. Unfortunately, a problem has emerged in this respect that we cannot estimate safely which component behaves well on a particular dataset without huge experimental work. Consequently, designers and developers must capture as many methods as possible during experimental work to establish which one is more appropriate for the specific problem. To solve this challenge, researchers have proposed customized classification pipelines based on a framework of various search algorithms, machine learning tools, and appropriate parameters for these algorithms that are capable of working independently of user knowledge. Until recently, the majority of these pipelines were constructed using genetic programming. In this paper, a new method is proposed for evolving classification pipelines automatically, founded on stochastic nature-inspired population-based optimization algorithms. The algorithms act as a tool for modeling customized classification pipelines consisting of the following tasks: choosing the proper preprocessing method, selecting the appropriate classification tool, and optimizing the model hyperparameters. The evaluation of the customized classification pipelines also showed potential for using the proposed method in the real world.