Light-weight algorithms for classification of time series

During this project, you will work at the intersection of three areas:

• Time series classification;
• Temporal convolutional networks;
• Hyperdimensional computing.

The project focuses on the task of time series classification, which involves assigning class labels to entire sequences of data points. Unlike traditional classification problems, time series classification requires models to account for the inherent ordering of input features. Many state-of-the-art machine learning algorithms do not fully account for temporal order. There are, however, recent developments [1] that try to address this issue by integrating temporal order into a high-performing family of algorithms using temporal convolutions. The integration is based on the ideas from hyperdimensional computing. The algorithm from [1] has been evaluated on the well-known UCR time series benchmark, where it demonstrated improvements on several datasets.

Many UCR datasets, however, do not require temporal order to successfully solve the classification problem, which limits the scope of the evaluation. A recent study [2] tried to address this limitation by introducing a new benchmark that is designed specifically to assess the temporal sensitivity. The goal of this project is to explore the applicability of machine learning algorithms (including the one from [1]) to this newly proposed time-sensitive benchmark [2] and to conduct a thorough evaluation of their performance.

 Key objectives

• Understand the principles of temporal convolutional networks, hyperdimensional computing, the algorithm from [1], and several other selected machine learning algorithms for time series classification.
• Implement and apply the algorithm from [1] and selected baseline algorithms to the recently proposed benchmark [2] designed to evaluate temporal sensitivity in time series classification.
• Evaluate the classification performance and complexity of the implemented algorithms to assess the benefits of incorporating temporal order in models for time series classification.
• Analyze results across datasets to identify when temporal sensitivity improves classification performance and when it does not.

Supervisors

Dr. Denis Kleyko

denis.kleyko@oru.se

Dr. Hadi Banaee

hadi.banaee@oru.se

[1] K. Schlegel, P. Neubert, and P. Protzel, “HDC-MiniROCKET: Explicit Time Encoding in Time Series Classification with Hyperdimensional Computing,” in International Joint Conference on Neural Networks (IJCNN), 2022, pp. 1–8. DOI: 10.1109/IJCNN55064.2022.9892158.

[2] Y. Zhang, G. Batista, and S. S. Kanhere, “Revisit Time Series Classification Benchmark: The Impact of Temporal Information for Classification,” in Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), 2025, pp. 193-205. DOI: 10.1007/978-981-96-8183-9_16.