Noise filter with hyperparameter recommendation: a meta-learning approach
Pedro B. Pio, Adriano Rívolli, André C. P. L. F. de Carvalho, Luís P. F. Garcia · 2023
Applying Machine Learning (ML) algorithms to a dataset can be time-consuming. It usually involves, not only selecting and fine-tuning the algorithm, but also other steps, such as data preprocessing. To reduce this time, the whole or a subset of this process has been automated by Automated ML (AutoML) techniques, which can include Bayesian Optimization, Genetic Programming, and Meta-Learning techniques. However, despite it often being a necessary stage, preprocessing is commonly not well handled in AutoML tools. In this work, we propose and experimentally investigate the use of meta-learning to recommend noise detection algorithms and the values for their hyperparameters. The proposed approach produces a ranking of the best noise filters for a given dataset, reducing the development cost of ML-based solutions and improving their predictive performance. To validate the process, we generated 10740 noisy datasets, which we describe using 97 meta-features. For each dataset, we applied 8 noise filters, which increased to 27 when we added variations of hyperparameter values. Next, we applied 4 ML algorithms to this data and created a performance ranking, which we used as a meta-target to induce 3 meta-regressors. We compared these 3 meta-regressors and the results with and without hyperparameters for the noise filters. According to the experimental results, the introduction of hyperparameter recommendation resulted in a higher gain in the F1-Score performance metric. However, it came at the cost of lower accuracy in the Top-K ranking evaluation.