Towards Similarity-Aware Time-Series Classification (2201.01413v2)

Abstract: We study time-series classification (TSC), a fundamental task of time-series data mining. Prior work has approached TSC from two major directions: (1) similarity-based methods that classify time-series based on the nearest neighbors, and (2) deep learning models that directly learn the representations for classification in a data-driven manner. Motivated by the different working mechanisms within these two research lines, we aim to connect them in such a way as to jointly model time-series similarities and learn the representations. This is a challenging task because it is unclear how we should efficiently leverage similarity information. To tackle the challenge, we propose Similarity-Aware Time-Series Classification (SimTSC), a conceptually simple and general framework that models similarity information with graph neural networks (GNNs). Specifically, we formulate TSC as a node classification problem in graphs, where the nodes correspond to time-series, and the links correspond to pair-wise similarities. We further design a graph construction strategy and a batch training algorithm with negative sampling to improve training efficiency. We instantiate SimTSC with ResNet as the backbone and Dynamic Time Warping (DTW) as the similarity measure. Extensive experiments on the full UCR datasets and several multivariate datasets demonstrate the effectiveness of incorporating similarity information into deep learning models in both supervised and semi-supervised settings. Our code is available at https://github.com/daochenzha/SimTSC

• The paper introduces SimTSC, a novel framework that blends similarity measures with deep learning to boost time-series classification accuracy.
• It leverages Graph Neural Networks and Dynamic Time Warping to model relationships among time-series data as nodes in a graph.
• Experiments demonstrate that SimTSC outperforms traditional methods in both supervised and semi-supervised settings, especially when labeled data is scarce.

Towards Similarity-Aware Time-Series Classification

The paper "Towards Similarity-Aware Time-Series Classification" addresses the challenge of improving Time-Series Classification (TSC) by integrating similarity-based methods with deep learning approaches. TSC is a fundamental task in time-series data mining with applications ranging from human activity recognition to healthcare and cybersecurity. Traditional approaches to TSC are generally categorized into two: similarity-based methods and deep learning models. The authors present a novel framework, Similarity-Aware Time-Series Classification (SimTSC), which synthesizes the strengths of both approaches to enhance classification performance, especially under limited supervision.

Key Contributions and Methodology

1. Integration of Similarity Measures and Deep Learning: The paper proposes SimTSC, which models similarity information using Graph Neural Networks (GNNs). This approach treats each time series as a node and pair-wise similarities as edges in a graph, thereby reformulating TSC as a node classification problem.
2. Graph Construction and Efficient Training: To leverage similarity information, the authors design an unsupervised graph construction strategy. They employ Dynamic Time Warping (DTW) as the similarity measure and utilize ResNet as the backbone neural architecture. Additionally, a batch training algorithm with negative sampling is introduced to improve efficiency, allowing SimTSC to scale to larger datasets.
3. Extensive Evaluations: The effectiveness of SimTSC is demonstrated through extensive experiments on the full UCR Time Series Classification Archive and a suite of multivariate datasets. The framework shows substantial improvements over baseline methods in both supervised and semi-supervised settings, notably in scenarios with limited labeled data.

Experimental Results

The authors conduct rigorous experiments comparing SimTSC against existing algorithms such as MLP, FCN, ResNet, InceptionTime, and TapNet. The results reveal that SimTSC, particularly in its semi-supervised variants, consistently achieves superior performance, especially when the number of training labels is limited. This indicates that the proposed integration of similarity measures into neural network training effectively aids in model generalization. Moreover, the paper provides insights into hyperparameter choices (e.g., scaling factor and number of neighbors), further optimizing SimTSC's performance.

Implications and Future Directions

SimTSC's capability to leverage both labeled and unlabeled data suggests significant practical implications for domains with abundant unlabeled time-series data but sparse labels. The integration of GNNs introduces a new dimension for exploring temporal dependencies, which may inspire further research into hybrid models that utilize structural properties of data.

For theoretical contributions, the demonstration of using GNNs in modeling similarity within time-series offers potential expansions into network-based reasoning for various data types. Future work may explore differentiable similarity measures such as soft DTW to provide end-to-end trainability and improve efficiency further. Additionally, extending SimTSC's application to other complex data domains, such as spatio-temporal datasets, presents a promising research avenue.

In conclusion, the paper provides valuable insights and a robust framework for advancing the state of TSC by integrating conventional similarity measures with modern deep learning techniques, enhancing model capability in diverse settings.