[論文レビュー] Deep Learning for Time Series Classification and Extrinsic Regression: A Current Survey

Navid Mohammadi Foumani, Lynn Miller+4 more|arXiv (Cornell University)|Feb 6, 2023

Time Series Analysis and Forecasting被引用数 10

ひとこと要約

A comprehensive survey of deep learning methods for time series classification (TSC) and extrinsic regression (TSER), covering architectures, self-supervised learning, data augmentation, transfer learning, and key applications like human activity recognition and Earth observation.

ABSTRACT

Time Series Classification and Extrinsic Regression are important and challenging machine learning tasks. Deep learning has revolutionized natural language processing and computer vision and holds great promise in other fields such as time series analysis where the relevant features must often be abstracted from the raw data but are not known a priori. This paper surveys the current state of the art in the fast-moving field of deep learning for time series classification and extrinsic regression. We review different network architectures and training methods used for these tasks and discuss the challenges and opportunities when applying deep learning to time series data. We also summarize two critical applications of time series classification and extrinsic regression, human activity recognition and satellite earth observation.

研究の動機と目的

• Summarize the current state-of-the-art in deep learning for TSC and TSER.
• Provide a taxonomy of deep learning methods and architectures used for time series data.
• Discuss training strategies, including self-supervised learning, data augmentation, and transfer learning.
• Highlight leading applications such as Human Activity Recognition and Satellite Earth Observation.
• Identify challenges and opportunities to guide future research.

提案手法

• Survey and synthesis of literature on deep learning for TSC/TSER.
• Taxonomy development from network configurations (MLP, CNN, RNN, GNN, Attention) and learning paradigms (supervised, self-supervised).
• Discussion of data augmentation and transfer learning strategies for time series.
• Review of representative architectures (InceptionTime, ResNet, FCN, Dilated CNNs, Disjoint-CNN, imaging-time-series approaches).
• Examination of self-supervised pretexts (contrastive learning, self-prediction) and attention/transformer variants for long-range dependencies.
• Description of two key applications: Human Activity Recognition and Earth Observation.

実験結果

リサーチクエスチョン

• RQ1 What are the main deep learning architectures and training paradigms currently used for TSC and TSER?
• RQ2 How do self-supervised learning and data augmentation contribute to performance in time series tasks?
• RQ3 What transfer learning strategies are effective for time series data?
• RQ4 What are the prominent applications driving advances in TSC/TSER, and what challenges remain?

主な発見

• Deep learning architectures (CNNs, RNNs, MLPs-adaptations, GNNs, attention-based models) dominate current TSC/TSER research and practice.
• Self-supervised learning and contrastive/self-prediction pretexts are emerging as promising for low-label regimes.
• Imaging-based transformations (GAF/MTF/GADF, RP, RPMCNN) provide alternative pathways to leverage 2D CNNs for time series.
• InceptionTime and related multi-scale/inception-based architectures often outperform older baselines on benchmark collections, with attention/transformer variants gaining interest for long-range dependencies.
• There is a recognized gap between traditional benchmarks (UCR/UEA) and deep learning data needs, motivating model and dataset co-design.
• Applications in Human Activity Recognition and Earth Observation are highlighted as key use cases driving methodology.