Meta-learning for the selection of algorithms and features for time series classification

Abstract

Due to the increasing collection of overtime observed data, time series are becoming a ubiquitous type of data in human beings' daily lives. This fact has also caused a significant increase in the number of techniques for time series classification in recent decades. In this scenario, when a researcher or developer needs a solution based on a classification model, there is a wide range of algorithms options to apply. When only one algorithm is chosen, its performance may not achieve satisfactory results for the problem to be solved. Even sets-based methods, which attempt to reduce the impacts of a bad algorithm choice, may fail on specific data sets. Therefore, the safest alternative is to perform cross-validation procedures to guide the classification algorithm choice. This approach usually guarantees a good choice, but it is very computationally expensive. In this context, we propose the creation of meta-learning-based recommendation models for time series classification algorithms. The idea behind this proposal is to describe previously analyzed problems (datasets) in order to efficiently induce meta-models through Machine Learning algorithms. These meta-models will be used to recommend classification algorithms for new datasets in order to avoid a very costly selection procedure. The Meta-learning will be evaluated at two levels: selection of a single classification algorithm and selection of parameters to compose an algorithm based on feature extraction through random convolutional kernels.(AU)