The CCA Missing Data Completion Algorithm for Machine Learning Purposes

Abstract The paper deals with multivariate time series classification, a widely used concept of supervised machine learning for automation in data analysis. Many algorithms have been proposed to deal with such classification, yet a majority of them require equal length data inputs, which often cannot be primarily satisfied. For this purpose different techniques of data adjustment are applied, such as truncation, padding or forecasting. In this paper we present a novel approach to data recovery, which yields an algorithm allowing to align variable-length time series, basing on cross-correlation. The paper presents the proposed algorithm architecture with its advantages and possible tasks to be adapted for. The novel algorithm is illustrated on a use case, which is a sequence of car-like vehicles driving through different road conditions. Two independent analyses using a specific time series classifier are derived to test it. The first one covers validation on the open-access data, while the second one focuses on verification on experimental data collected by ourselves. In both cases the incomplete data is recovered with the use of a few chosen techniques including the one proposed in this paper. The results are thereafter analysed and compared using fundamental statistical factors.