Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ECRTime: Ensemble Integration of Classification and Retrieval for Time Series Classification

Fan Zhao, You Chen

TL;DR

This paper tackles time series classification by addressing limitations of the common FC+SoftMax paradigm, notably inter-class similarity and intra-class inconsistency seen in UCR datasets. It introduces ECR, a dual-branch framework that jointly learns classification and retrieval representations, and then extends this to ECRTime, an ensemble of three ECRs that achieves state-of-the-art performance among deep-learning TSC methods while reducing training time relative to InceptionTime. By replacing SoftMax with a 1-NN-style retrieval objective and aligning the training loss with retrieval-based objectives, the authors demonstrate improved accuracy across 112 UCR datasets, with ECRTime further boosting results through model ensembling. The approach balances accuracy and efficiency, offering practical benefits for large-scale TSC tasks and suggesting future work on retrieval-enabled TSC and multi-dimensional time series.

Abstract

Deep learning-based methods for Time Series Classification (TSC) typically utilize deep networks to extract features, which are then processed through a combination of a Fully Connected (FC) layer and a SoftMax function. However, we have observed the phenomenon of inter-class similarity and intra-class inconsistency in the datasets from the UCR archive and further analyzed how this phenomenon adversely affects the "FC+SoftMax" paradigm. To address the issue, we introduce ECR, which, for the first time to our knowledge, applies deep learning-based retrieval algorithm to the TSC problem and integrates classification and retrieval models. Experimental results on 112 UCR datasets demonstrate that ECR is state-of-the-art(sota) compared to existing deep learning-based methods. Furthermore, we have developed a more precise classifier, ECRTime, which is an ensemble of ECR. ECRTime surpasses the currently most accurate deep learning classifier, InceptionTime, in terms of accuracy, achieving this with reduced training time and comparable scalability.

ECRTime: Ensemble Integration of Classification and Retrieval for Time Series Classification

TL;DR

This paper tackles time series classification by addressing limitations of the common FC+SoftMax paradigm, notably inter-class similarity and intra-class inconsistency seen in UCR datasets. It introduces ECR, a dual-branch framework that jointly learns classification and retrieval representations, and then extends this to ECRTime, an ensemble of three ECRs that achieves state-of-the-art performance among deep-learning TSC methods while reducing training time relative to InceptionTime. By replacing SoftMax with a 1-NN-style retrieval objective and aligning the training loss with retrieval-based objectives, the authors demonstrate improved accuracy across 112 UCR datasets, with ECRTime further boosting results through model ensembling. The approach balances accuracy and efficiency, offering practical benefits for large-scale TSC tasks and suggesting future work on retrieval-enabled TSC and multi-dimensional time series.

Abstract

Deep learning-based methods for Time Series Classification (TSC) typically utilize deep networks to extract features, which are then processed through a combination of a Fully Connected (FC) layer and a SoftMax function. However, we have observed the phenomenon of inter-class similarity and intra-class inconsistency in the datasets from the UCR archive and further analyzed how this phenomenon adversely affects the "FC+SoftMax" paradigm. To address the issue, we introduce ECR, which, for the first time to our knowledge, applies deep learning-based retrieval algorithm to the TSC problem and integrates classification and retrieval models. Experimental results on 112 UCR datasets demonstrate that ECR is state-of-the-art(sota) compared to existing deep learning-based methods. Furthermore, we have developed a more precise classifier, ECRTime, which is an ensemble of ECR. ECRTime surpasses the currently most accurate deep learning classifier, InceptionTime, in terms of accuracy, achieving this with reduced training time and comparable scalability.
Paper Structure (22 sections, 6 equations, 19 figures, 2 tables)

This paper contains 22 sections, 6 equations, 19 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. State-Of-The-Art Time Series Classifiers
  4. Deep learning-based methods
  5. Method
  6. ECR Framework
  7. Backbone network structure
  8. Loss
  9. Distance
  10. Ensemble
  11. Experiments and results
  12. Experiment setup
  13. Comparing with deep learning-based methods
  14. Comparing with SOTAs
  15. Runtime analysis
  16. ...and 7 more sections

Figures (19)

  • Figure 1: Critical difference diagram showing the performance of ECRTime compared to the current SOTA classifiers on 112 datasets from the UCR archive.
  • Figure 2: Visualization of the CinCECGTorso dataset from the UCR archive, which comprises four categories. The display showcases the first 10 sequences sampled from each category.
  • Figure 3: Conducting a heatmap analysis based on DTW distance for series of class1 and class2 in the \ref{['CinCECGTorso_numclass4']}.
  • Figure 4: Within the framework of the "FC+SoftMax" classification paradigm, the distribution of feature vectors for the CinCECGTorso dataset is visualized using t-SNE. The four colors in the figure represent four classes. The left graph represents the distribution of features from the test set and proxy features, while the right graph illustrates the distribution of features from the training set (used as a reference library) and features from the test set.
  • Figure 5: Training pipeline(top) and inference pipeline(bottom) of the ECR. GAP: Global Average Pooling, FC: Fully Connected Layer, L2Norm: L2 normalization. In the top subplot, the blue and red boxes represent time series of different categories.
  • ...and 14 more figures