Abstract Deep learning has achieved remarkable success in various domains including time series analysis, computer vision and natural language processing. However, high computational and memory demands of state-of-the-art architectures pose challenges for deployment in resource-limited environments. Knowledge Distillation (KD) addresses this by transferring knowledge from a large teacher model to a smaller, more efficient student model while maintaining competitive performance. In this work, we investigate the effectiveness of KD for Time Series Classification (TSC) across three architectures: the classical Fully Convolutional Network (FCN), the convolutional Inception model and the transformer-based ConvTran model. We evaluate our approach on UCR Archive, the largest benchmark repository of time series datasets, by modifying architectural components such as convolutional filters, Inception modules and attention heads across the three architectures. Our results consistently show that KD most effectively benefits student models of intermediate complexity across all three architectures, with the distilled FCN student reducing parameters by a factor of 38, the distilled Inception student achieving nearly the same performance as the teacher with 42% fewer parameters and the distilled ConvTran student with 2 attention heads showing the most significant improvement through distillation. To encourage further research and reproducibility, we provide our implementation at https://github.com/MSD-IRIMAS/KD-4-TSC.

Time series classification has achieved significant advancements through deep learning models; however, these models often suffer from high complexity and computational costs. To address these challenges while maintaining effectiveness, we introduce COCALITE, an innovative hybrid model that combines the efficient LITE model with an augmented version incorporating Catch22 features during training. COCALITE operates with only 4.7% of the parameters of the state-of-the-art Inception model, significantly reducing computational overhead. By integrating these complementary approaches, COCALITE leverages both effective feature engineering and deep learning techniques to enhance classification accuracy. Our extensive evaluation across 128 datasets from the UCR archive demonstrates that COCALITE achieves competitive performance, offering a compelling solution for resource-constrained environments.

Time series classification and regression are rapidly evolving fields that find areas of application in all domains of machine learning and data science. This hands on tutorial will provide an accessible overview of the recent research in these fields, using code examples to introduce the process of implementing and evaluating an estimator. We will show how to easily reproduce published results and how to compare a new algorithm to state-of-the-art. Finally, we will work through real world examples from the field of Electroencephalogram (EEG) classification and regression. EEG machine learning tasks arise in medicine, brain-computer interface research and psychology. We use these problems to how to compare algorithms on problems from a single domain and how to deal with data with different characteristics, such as missing values, unequal length and high dimensionality. The latest advances in the fields of time series classification and regression are all available through the aeon toolkit, an open source, scikit-learn compatible framework for time series machine learning which we use to provide our code examples.

Deep learning models have been shown to be a powerful solution for Time Series Classification (TSC). State-of-the-art architectures, while conducting promising results on the UCR archive, present a high number of trainable parameters. This can lead to long training with a high CO2, Power consumption and possible increase in the number of FLoat-point Operation Per Second (FLOPS). In this paper, we present a new architecture for TSC, the Light Inception with boosTing tEchnique (LITE) with only 2.34% of the state-of-the-art model InceptionTime’s number of parameters, while preserving performance. This architecture, with only 9, 814 trainable parameters due to the usage of DepthWise Separable Convolutions (DWSC), is boosted by three techniques: multiplexing, custom filters, and dilated convolution. The LITE architecture, trained on the UCR, is 2.78 times faster than InceptionTime and consumes 2.79 times less CO2 and Power.

In recent years, there has been an increasing interest in Deep Learning models for time series classification. In this field, state-of-the-art architectures rely on convolution neural networks that learn one dimensional filters in order to capture patterns allowing to discriminate between the different classes. These filters are randomly initialized and modified throughout model training. In this paper, we explore the creation of handcrafted (non learned) filters in order to capture specific patterns in a time series. We propose a set of filters whose values are fixed and not modified during the training step. Our goal with these filters is to captures pecific patterns in a time series (increase, decrease, peaks) and study the relevance of adding such filters to existing architectures ranging from simple architecture (Fully Convolutional Network (FNC)) to state-of-the-art architecture (InceptionTime). Experiments reveal that adding our manually created filters increase the prediction accuracy on a majority of the 128 datasets of the UCR Archive. They also show that handcrafted filters and learned filters are complementary to obtain the best preforming models. This work is the first step in proposing a catalog of generic and fixed filters that could be useful in a large range of applications to improve deep models accuracy for time series classification.

Transfer learning for deep neural networks is the process of first training a base network on a source dataset, and then transferring the learned features (the network’s weights) to a second network to be trained on a target dataset. This idea has been shown to improve deep neural network’s generalization capabilities in many computer vision tasks such as image recognition and object localization. Apart from these applications, deep Convolutional Neural Networks (CNNs) have also recently gained popularity in the Time Series Classification (TSC) community. However, unlike for image recognition problems, transfer learning techniques have not yet been investigated thoroughly for the TSC task. This is surprising as the accuracy of deep learning models for TSC could potentially be improved if the model is fine-tuned from a pre-trained neural network instead of training it from scratch. In this paper, we fill this gap by investigating how to transfer deep CNNs for the TSC task. To evaluate the potential of transfer learning, we performed extensive experiments using the UCR archive which is the largest publicly available TSC benchmark containing 85 datasets. For each dataset in the archive, we pre-trained a model and then fine-tuned it on the other datasets resulting in 7140 different deep neural networks. These experiments revealed that transfer learning can improve or degrade the models predictions depending on the dataset used for transfer. Therefore, in an effort to predict the best source dataset for a given target dataset, we propose a new method relying on Dynamic Time Warping to measure inter-datasets similarities. We describe how our method can guide the transfer to choose the best source dataset leading to an improvement in accuracy on 71 out of 85 datasets.

Time series classification maps time series to labels. The nearest neighbor algorithm (NN) using the Dynamic Time Warping (DTW) similarity measure is a leading algorithm for this task and a component of the current best ensemble classifiers for time series. However, NN-DTW is only a winning combination when its meta-parameter – its warping window – is learned from the training data. The warping window (WW) intuitively controls the amount of distortion allowed when comparing a pair of time series. With a training database of N time series of lengths L, a naive approach to learning the WW requires Θ(N2·L3) operations. This often results in NN-DTW requiring days for training on datasets containing a few thousand time series only. In this paper, we introduce FastWWSearch: an efficient and exact method to learn WW. We show on 86 datasets that our method is always faster than the state of the art, with at least one order of magnitude and up to 1000x speed-up.

While the Dynamic Time Warping (DTW) — based Nearest-Neighbor Classification algorithm is regarded as a strong baseline for time series classification, in recent years there has been a plethora of algorithms that have claimed to be able to improve upon its accuracy in the general case. Many of these proposed ideas sacrifice the simplicity of implementation that DTW-based classifiers offer for rather modest gains. Nevertheless, there are clearly times when even a small improvement could make a large difference in an important medical or financial domain. In this work, we make an unexpected claim; an underappreciated “low hanging fruit” in optimizing DTW's performance can produce improvements that make it an even stronger baseline, closing most or all the improvement gap of the more sophisticated methods. We show that the method currently used to learn DTW's only parameter, the maximum amount of warping allowed, is likely to give the wrong answer for small training sets. We introduce a simple method to mitigate the small training set issue by creating synthetic exemplars to help learn the parameter. We evaluate our ideas on the UCR Time Series Archive and a case study in fall classification, and demonstrate that our algorithm produces significant improvement in classification accuracy.

In machine learning, data augmentation is the process of creating synthetic examples in order to augment a dataset used to learn a model. One motivation for data augmentation is to reduce the variance of a classifier, thereby reducing error. In this paper, we propose new data augmentation techniques specifically designed for time series classification, where the space in which they are embedded is induced by Dynamic Time Warping (DTW). The main idea of our approach is to average a set of time series and use the average time series as a new synthetic example. The proposed methods rely on an extension of DTW Barycentric Averaging (DBA), the averaging technique that is specifically developed for DTW. In this paper, we extend DBA to be able to calculate a weighted average of time series under DTW. In this case, instead of each time series contributing equally to the final average, some can contribute more than others. This extension allows us to generate an infinite number of new examples from any set of given time series. To this end, we propose three methods that choose the weights associated to the time series of the dataset. We carry out experiments on the 85 datasets of the UCR archive and demonstrate that our method is particularly useful when the number of available examples is limited (e.g. 2 to 6 examples per class) using a 1-NN DTW classifier. Furthermore, we show that augmenting full datasets is beneficial in most cases, as we observed an increase of accuracy on 56 datasets, no effect on 7 and a slight decrease on only 22.

Recent years have seen significant progress in improving both the efficiency and effectiveness of time series classification. However, because the best solution is typically the Nearest Neighbor algorithm with the relatively expensive Dynamic Time Warping as the distance measure, successful deployments on resource constrained devices remain elusive. Moreover, the recent explosion of interest in wearable devices, which typically have limited computational resources, has created a growing need for very efficient classification algorithms. A commonly used technique to glean the benefits of the Nearest Neighbor algorithm, without inheriting its undesirable time complexity, is to use the Nearest Centroid algorithm. However, because of the unique properties of (most) time series data, the centroid typically does not resemble any of the instances, an unintuitive and underappreciated fact. In this work we show that we can exploit a recent result to allow meaningful averaging of 'warped' times series, and that this result allows us to create ultra-efficient Nearest 'Centroid' classifiers that are at least as accurate as their more lethargic Nearest Neighbor cousins.