Synchronized Time Profile Similarity in Applications to Nearest Neighbor Classification

Abstract

One of the existing approaches to time series classification exploits the time profiles using the original data with synchronization instead of model-implied data. Synchronization aligns inter-individual data from different time points to account for potential phase offsets and nonstationarity in the data. Such synchronization has been applied in psychology: For example, coordinated motion between two individuals exchanging information was used as a predictor and outcome of psychological processes. Synchronization also affords better classification outcomes, as discussed in the data mining community, through aligning the data to reveal the maximally shared profile underlying two compared data sequences. For inter-individual comparison of univariate time series data, existing similarity indices include Euclidean distances and squared correlations. For synchronization, we introduce dynamic time warping and window-crossed lagging. The current study compares the Euclidean distance and the squared correlation before and after synchronization using window-crossed lagging and dynamic time warping in applications to one-nearest-neighbor classification tasks. Discussion, limitations, and future directions are provided.