Robust Unsupervised Feature Learning from Time-Series
Jianyu Miao, Yong Shi, Lingfeng Niu · 2016
Feature learning from unlabeled times series data is an important component in data analysis. Shaplets are discriminative sub-sequence of time series that can best predict target variable. Therefore, shaplets discovery is very important for analysis of time-series. Recently, based on optimization model, a novel approach has been proposed to learning shaplets. To make shaplets learning model more robust, in this paper, we propose a new unsupervised shaplets learning model with emphasizing joint l2,1norm minimization on both loss function and regularization, which can efficiently degenerate effect of noisy and outliers. The utilization of constraints can make the pseudo labels learned more accurate. Numeral experiments validate the effectiveness of the proposed method.