Semi-Supervised TSC by Temporal Relation Prediction


Contents

  1. Abstract
  2. Introduction
  3. Method
    1. Training on labeled data
    2. Training on unlabeled data


0. Abstract

Few efforts consider the underlying temporal relation structure of TS

\(\rightarrow\) propose SemiTime

( = a simple and effective method of Semi-supervised TSC )


For LABELED TS …

  • conducts the supervised cls


For UNLABELED TS …

  • the segments of past future pair are sampled from TS

  • 2 segments of pair from the same TS = positive

    ( \(\leftrightarrow\) negative )

  • temporal relation between those segments is predicted by SemiTime


By jointly (1) classifying labeled data & (2) predicting the temporal relation of unlabeled data

\(\rightarrow\) useful representation of unlabeled TS can be captured by SemiTime


1. Introduction

underlying temporal relation of TS is a significant supervision signal

propose a general semi-supervised TSC

  • by exploring the semantic feature from unlabeled data

figure2


2. Method

figure2


proposed ”SemiTime” consists of 3 modules

  • (1) temporal relational segment sampling module
  • (2) supervised classification module
  • (3) self-supervised temporal relation prediction module


Input : \(\left(\boldsymbol{t}_i, y_i\right) \in \mathcal{D}_L\) & \(\boldsymbol{t}_i \in \mathcal{D}_U\)

  • where \(\mathcal{D}_U=\left\{\boldsymbol{t}_i \mid \boldsymbol{t}_i=\left(t_{(i, 1)}, \ldots t_{(i, T)}\right)\right\}_{i=1}^N\) : set of \(T\)-length TS
  • where \(\mathcal{D}_L\) is subset of \(\mathcal{D}_U\)


Notation

  • backbone encoder : \(f_\theta\)
  • classification head : \(h_\mu\)
  • relation head : \(h_{\varphi}\)


(1) Training on labeled data

Input : \(\left(\boldsymbol{t}_i, y_i\right) \in \mathcal{D}_L\)


Representation : \(\boldsymbol{z}_i=f_\theta\left(\boldsymbol{t}_i\right)\)


CLS output : \(p_i=h_\mu\left(\boldsymbol{z}_i\right)\)


Loss : \(\mathcal{L}_{c l s}=-\frac{1}{ \mid \mathcal{D}_L \mid } \sum_{i=1}^{ \mid \mathcal{D}_L \mid } y_i \cdot \log \left(p_i\right)\).

  • CE loss


(2) Training on unlabeled data

Input : \(\boldsymbol{t}_i \in \mathcal{D}_U\)


Split input into two parts

  • (1) front \(B\)-length part of \(\boldsymbol{t}_i\) : past segment \(\boldsymbol{s}_{i, \alpha}\)
  • (2) rear \((T-B)\)-length part of \(\boldsymbol{t}_i\) : future segment \(\boldsymbol{s}_{i, \alpha}^{+}\)
    • where \(B=\lfloor\alpha * T\rfloor\) and \(\alpha\) is a past-future segment split ratio


Anchor & Pos & Neg

  • Anchor : \(\boldsymbol{s}_{i, \alpha}\)
  • Pos : \(\boldsymbol{s}_{i, \alpha}^{+}\) ( from same TS \(\boldsymbol{t}_i\) )
  • Neg : \(s_{j, \alpha}^{-}\)


Representation :

  • \(\boldsymbol{z}_{i, \alpha}=f_\theta\left(\boldsymbol{s}_{i, \alpha}\right)\).
  • \(\boldsymbol{z}_{i, \alpha}^{+}=f_\theta\left(\boldsymbol{s}_{i, \alpha}^{+}\right)\).
  • \(\boldsymbol{z}_{j, \alpha}^{-}=f_\theta\left(s_{i, \alpha}^{-}\right)\).


CLS output :

  • \(p_{2 i-1}=h_{\varphi}\left(\left[\boldsymbol{z}_{i, \alpha}, \boldsymbol{z}_{i, \alpha}^{+}\right]\right)\)……… POS relation prediction
  • \(p_{2 i}=h_{\varphi}\left(\left[\boldsymbol{z}_{i, \alpha}, \boldsymbol{z}_{i, \alpha}^{-}\right]\right)\)……… NEG relation prediction


Loss : \(\mathcal{L}_{r e l}=-\frac{1}{2 \mid \mathcal{D}_U \mid } \sum_{i=1}^{2 \mid \mathcal{D}_U \mid } \tilde{y}_i \cdot \log \left(p_i\right)+\left(1-\tilde{y}_i\right) \cdot\left(1-\log \left(p_i\right)\right)\)

  • binary CE loss
  • where \(\tilde{y}_i=1\) denotes positive relation and \(\tilde{y}_i=0\) negative relation


figure2

Categories: ,

Updated: