Visualizing Time Series Data with Temporal Matching Based t-SNE

Interpreting time series data has always been a hot research topic for various applications, especially when the dimensionality of time series dataset keeps growing almost prohibitively as technology advances. There exist several dimensionality reduction techniques attempting to address the related problems, but the inherent nature of time series datasets usually involves factors including time and amplitude shifting and scaling, which could impact the trustworthiness of the visualization results. t-distributed Stochastic Neighbor Embedding (t-SNE) is considered as a highly effective machine learning algorithm for visualization and it is in fact a nonlinear dimensionality reduction technique tailored made to embed high-dimensional data in a low-dimensional space of two to three dimensions only for proper visualization. In view of the key problem of adopting t-SNE to visualize time series data, we propose to introduce two temporal matching metrics, namely, dynamic time warping (DTW) and angular metric for shape similarity (AMSS), for t-SNE to enhance its time series data visualization ability. They provide a more robust similarity metric for time series data so that the embedding process in t-SNE can be made more effective, as demonstrated by various data visualization experiments.