Transfer Learning with Time Series Data : A Systemetic Mapping Study

Contents

  1. Abstract
  2. Overview & Definitions
    1. Time Series (TS)
    2. TS problems
    3. Transfer Learning
    4. TL solution approachs


0. Abstract

Transfer Learning (TL) :

  • relaxes the assumption that train & test data need to be drawn from same distn
  • benefits from various TS domains

conduct a study of literature on TL with TS data


1. Overview & Definitions

(1) Time Series (TS)

  • Time Series : \(T=\left[x_1, \ldots, x_n\right]\)
    • data points \(x_i\) of length \(n\)
  • Uni & Multi-variate TS
    • Univariate : TS where \(x_i \in \mathbb{R}\).
    • Multivariate : TS where each \(x_i\) is a d-dim vector of real values \(\left(x_i^1, \ldots, x_i^d\right), x_i^j \in \mathbb{R}\).


(2) TS problems

  1. TS classification

    • Assign a TS ( or subsequence TS ) a class \(c_i\), out of \(C=\left\{c_1, \ldots, c_n \mid n \geq 2\right\}\)
  2. TS regression

    • For TS \(T\), predict numeric value (\(y\)) / values (\(y_1, \ldots, y_n\))
  3. TS clustering

    • Assign a TS ( or subsequence TS ) a cluster \(c_i\), out of \(C=\left\{c_1\right. \left.\ldots, c_n \mid n \geq 1\right\}\)

      based on similarity measure \(\operatorname{Sim}(a, b)\)

  4. TS anomaly detection

    • Assign a TS ( or subsequence TS ) to one of \(\left\{c_{\text {normal }}, c_{\text {anomaly }}\right\}\)
  5. TS forecasting

    • Given \(T=\left[x_1, \ldots, x_n\right]\), predict..
      • (1) single-step : \(x_{n+1}\)
      • (2) multi-step : \(x_{n+1}, \ldots, x_{n+m}\)


(3) Transfer Learning

  • transfer knowledge from one domain to another simliar domain

  • to improve the generalization ability


Notation

  • (SOURCE) domain

    • source domain : \(D_S\)
    • source domain learning task : \(T_S\)
    • model : \(f_S(\cdot)\)
  • (TARGET) domain

    • target domain : \(D_T\)
    • target domain learning task : \(T_T\)
    • model : \(f_T(\cdot)\)
  • Goal : improve \(f_T(\cdot)\) using the knowledge in \(D_S\) & \(T_S\) ,

    where \(D_S \neq D_T\) or \(T_S \neq T_T\)


Types of TL

  • (1) Domain Adaptation : \(D_S \neq D_T\)
  • (2) Task Adaptation : \(T_S \neq T_T\)
  • Combination = (1) + (2)


figure2


(4) TL solution approachs

  1. instance-based

    • selection or reweighting of samples from the source domain
    • assumption ) instance from source domain are more/less similar to the set of target domain instances
  2. feature-representation-based ( = mapping-based )

    • map into common feature space

      ( place of features, representating characteristics of both domains )

  3. parameter-based ( = network transfer )

    • use pre-trained model
  4. relational-knowledge-based

    • not applicable to TS


figure2