A0753
Title: Using t-SNE in analyzing multivariate time series data
Authors: Soudeep Deb - Indian Institute of Management Bangalore (India) [presenting]
Abstract: In multivariate data, t-distributed stochastic neighbour embedding (t-SNE) is one of the advanced dimension reduction techniques. Albeit it is primarily used for visualization in lower dimensions, in the context of time series analysis, it has the potential to be utilized in other problems too. Two different applications of t-SNE in analyzing multivariate time series are discussed. First, a classification technique that uses the advantages of dimension reduction through t-SNE is proposed, coupled with the attractive properties of nonparametric spectral density estimates and the k-nearest neighbour technique. The theoretical consistency of the proposed algorithm is proved, and the efficacy of the method is shown using an interesting dataset from medical research. The second part is about a new methodology to detect structural breaks in multivariate time series. Once again, the same principles are used, and t-SNE with nonparametric spectral density estimates in lower dimensions is utilized. Relevant empirical justifications are obtained to demonstrate the accuracy of the proposed method in detecting structural breaks in multivariate series. For application, the exchange rates of the Indian Rupee against four other major currencies from the last decade are considered.
