Deep MTS Embedding Clustering via Attentive-Gated Autoencoder (2020)

Contents

  1. Abstract
  2. Introduction
  3. DeTSEC : Deep Time Series Embedding Clustering


0. Abstract

propose a DL-based framework for clustering MTS, with varying length

\(\rightarrow\) propose DeTSEC (Deep Time Series Embedding Clustering)


1. Introduction

DeTSEC (Deep Time Series Embedding Clustering)

  • different domains OK
  • varying length OK


2 stages

  • step 1) Recurrent autoencoder exploits attention & gating mechanism to produce a preliminary embedding representation
  • step 2) Clustering Refinement stage
    • stretch the embedding manifold towards the corresponding cluters


2. DeTSEC : Deep Time Series Embedding Clustering

Notation

  • \(X=\left\{X_{i}\right\}_{i=1}^{n}\) : multivariate time-series

    • \(X_{i} \in X\) : time-series

      where \(X_{i j} \in R^{d}\) = multidimensional vector of the time-series \(X_{i}\) at timestamp \(j\), with \(1 \leq j \leq T\)

    • \(d\) : dimensionality of \(X_{i j}\)

    • \(T\) : maximum time-series length

  • \(X\) can contain time-series with DIFFERENT length


Goal

  • partition \(X\) in a given number of clusters


2 stages

  • stage 1) GRU based autoencoder

    • for each GRU unit, attention is applied,

      to combine the information coming from different timestamps

  • stage 2) refine the representation, by taking into account a 2-fold task

    • 1) reconstruction
    • 2) another one devoted to stretch the embedding manifold towards clustering centroids


3 different compontents

  • 1) encoder
  • 2-1) backward decoder
  • 2-2) forward decoder


figure2


Loss Function

(1) autoencoder network

\(\begin{aligned} L_{a e}=& \frac{1}{ \mid X \mid } \sum_{i=1}^{ \mid X \mid } \mid \mid X_{i}-\operatorname{dec}\left(\operatorname{enc}\left(X_{i}, \Theta_{1}\right), \Theta_{2}\right) \mid \mid _{2}^{2} \\ &+\frac{1}{ \mid X \mid } \sum_{i=1}^{ \mid X \mid } \mid \mid \operatorname{rev}\left(X_{i}\right)-\operatorname{dec}_{b a c k}\left(\operatorname{enc}\left(X_{i}, \Theta_{1}\right), \Theta_{3}\right) \mid \mid _{2}^{2} \end{aligned}\).


(2) regularizer term

\(\frac{1}{ \mid X \mid } \sum_{i=1}^{ \mid X \mid } \sum_{l=1}^{n C l u s t} \delta_{i l} \mid \mid \text { Centroids }_{l}-\operatorname{enc}\left(X_{i}, \Theta_{1}\right) \mid \mid _{2}^{2}\).


(3) Total loss : (1) + (2)

Tags:

Categories:

Updated: