Abstract

This paper introduces NN-STNE, a neural network using t-distributedstochastic neighbor embedding (t-SNE) as a hidden layer to reduce inputdimensions by mapping long time-series data into shapelet membershipprobabilities. A Gaussian kernel-based mean square error preserves local datastructure, while K-means initializes shapelet candidates due to the non-convexoptimization challenge. Unlike existing methods, our approach uses t-SNE toaddress crowding in low-dimensional space and applies L1-norm regularization tooptimize shapelet length. Evaluations on the UCR dataset and an electricalcomponent manipulation task, like switching on, demonstrate improved clusteringaccuracy over state-of-the-art feature-learning methods in robotics.

 

Quick Read (beta)