Recurrence Plots and Cross Recurrence Plots

Knowledge-Based Systems, 309, 112864p. (2025) DOI:10.1016/j.knosys.2024.112864

Visemble: A fast ensemble approach for time series classification with multiple visual representations

Time series are prevalent data in finance, smart cities, sensor networks, engineering, bioinformatics, among other domains. These data differ from regular tabular data as they involve sequences of observations at successive, equally spaced points in time. The temporal ordering of observations in time series carries significant information, making pattern detection in subsequences crucial for supervised learning tasks such as forecasting and classification. Current classification methods often overlook the potential of transformations that represent time series as images. Visual representations such as Recurrence Plots and Markov Transition Fields capture intricate temporal dynamics and spatial patterns that traditional one-dimensional approaches might miss. We propose Visemble, an efficient and accurate ensemble method for time series classification that leverages multiple visual representations and weighted soft voting. Besides our proposal, we present a broader discussion regarding time series imaging representations. Visemble combines the original time domain representation, first-order differences, and various two-dimensional representations. Our ensemble integrates accurate and diverse classifiers to improve the classification performance of individual imaging based models efficiently using Random Forest. A comprehensive experimental evaluation with hundreds benchmark datasets from varying domains demonstrated that Visemble achieves comparable accuracy to state-of-the-art ensemble based methods, such as Elastic Ensemble, BOSS, and Temporal Dictionary Ensemble, but with significantly reduced computational time.