Imaging Time-Series to Improve Classification and Imputation (1506.00327v1)

Abstract: Inspired by recent successes of deep learning in computer vision, we propose a novel framework for encoding time series as different types of images, namely, Gramian Angular Summation/Difference Fields (GASF/GADF) and Markov Transition Fields (MTF). This enables the use of techniques from computer vision for time series classification and imputation. We used Tiled Convolutional Neural Networks (tiled CNNs) on 20 standard datasets to learn high-level features from the individual and compound GASF-GADF-MTF images. Our approaches achieve highly competitive results when compared to nine of the current best time series classification approaches. Inspired by the bijection property of GASF on 0/1 rescaled data, we train Denoised Auto-encoders (DA) on the GASF images of four standard and one synthesized compound dataset. The imputation MSE on test data is reduced by 12.18%-48.02% when compared to using the raw data. An analysis of the features and weights learned via tiled CNNs and DAs explains why the approaches work.

• The paper presents novel time series representations (GASF/GADF and MTF) that convert temporal data into images to capture intricate dependencies.
• The methodology employs Tiled CNNs and Denoised Auto-Encoders, achieving superior classification on 9 of 20 datasets and reducing imputation errors by up to 48%.
• The research demonstrates the potential of integrating computer vision with time series analysis, paving the way for improved real-world data processing applications.

Imaging Time-Series to Improve Classification and Imputation

The paper presents a robust framework for enhancing time series classification and imputation by employing imaging techniques inspired by deep learning advancements in computer vision. The authors propose novel methodologies for encoding time series data into visual representations: Gramian Angular Summation/Difference Fields (GASF/GADF) and Markov Transition Fields (MTF). These transformations facilitate the application of computer vision techniques to time series data, leveraging deep learning architectures such as Tiled Convolutional Neural Networks (Tiled CNNs) and Denoised Auto-encoders (DA).

Key Contributions

1. Novel Time Series Representations:
   • GASF/GADF: By converting time series data into a polar coordinate system, these methods capture temporal correlations, effectively encoding them in a compact image format that preserves temporal dependencies.
   • MTF: Encodes transition probabilities from Markov models, maintaining temporal structure through dynamic transition statistics.
2. Deep Learning Application:
   • Tiled CNNs extract meaningful features from the transformed images to classify data across 20 datasets. The model achieved competitive classification performance, surpassing nine established methods on several datasets.
   • Denoised Auto-encoders (DA), leveraging the bijection property of the GASF transformation, showed substantial improvements (12.18%-48.02% reduction in MSE) over raw data imputation techniques across various datasets.

Methodology

• Tiled CNNs: Utilizing Tiled CNNs allowed the authors to capture invariant features through tiled weight sharing and multiple feature maps. This architecture facilitates learning overcomplete representations that are critical for efficient and accurate classification.
• Denoised Auto-Encoders: For imputation, DAs were applied to recover missing time series data by predicting from the GASF images. The bijective nature of the GASF transformation permits precise recovery, demonstrating the potential of imaging-based approaches for time series imputation.

Numerical Results and Analysis

The paper reports strong numerical results, particularly in classification tasks, where the proposed GASF-GADF-MTF combination achieved the best performance on 9 out of 20 datasets. For imputation, the GASF method consistently outperformed raw data strategies, indicating the advantage of the temporal features encoded within the GASF images.

Implications and Future Research

The research opens paths for integrating visual analysis techniques into time series data processing, suggesting a paradigm shift in how such data can be handled computationally. Practically, this approach could be extended to real-world applications in fields requiring reliable time series classification and imputation. Theoretically, the capacity of deep learning to manage rich temporal information materializes through these innovative transformations.

Future work may explore recurrent neural networks for handling streaming data, expanding the utility of the GAF and MTF approaches. Extending the deep learning architecture to more sophisticated models could further enhance feature extraction and improve classification accuracy. Furthermore, a detailed exploration into the generative capabilities of these models might reveal more about their strengths in time series analysis.

The integration of visual encoding techniques into the time series domain signifies a noteworthy advancement in the field, marking a significant contribution to both the theoretical framework and practical applications of machine learning in temporal data processing.