Differential Correlation Approach for Multivariate Time Series Feature Selection

Abstract

In Machine Learning, feature selection is paramount to achieve accurate and correct results as well as to reduce the computational effort. As a first step, it can be used to identify erroneous features or features introducing spurious correlation. This also applies to machine learning in an engineering environment. Here, data is often given as multivariate time series datasets, which require adapted approaches for machine learning as well as preprocessing steps, like feature selection. Because of the importance of feature selection, special methods for this use case are needed. Therefore, we propose an applicable feature selection method for multivariate time series, based on a differential correlation approach. This method aims to detect erroneous features or those introducing spurious correlation through the assessment of variable relationships across multiple multivariate time series instances. For this, the correlation between the time series features is compared across all examined instances through a differential correlation approach. To show the applicability of the proposed selection method, it is evaluated on artificially generated time series datasets, including erroneous features. Additionally, the method is evaluated on a real multivariate time series dataset in context of an engineering environment.