EE Seminar: Lean and Early: Feature Selection for Fast Human Activity Recognition from Time Series Data
https://tau-ac-il.zoom.us/j/84357771271
Electrical Engineering Systems ZOOM Seminar
Speaker: Doron Almog
M.Sc. student under the supervision of Prof. Neta Rabin
Monday, 18th August 2025, at 15:00
Lean and Early: Feature Selection for Fast Human Activity Recognition from Time Series Data
Abstract
Time series classification involves assigning labels to sequences of time-ordered data. This task is essential in human activity recognition (HAR) for classifying hand and limb movements from signal data in applications such as prosthetics control.
This thesis tackles the challenge of processing high-dimensional feature representations generated from time series data in multiclass HAR tasks. The Random Convolutional Kernels (ROCKET) approach was employed; in this case, the lean version of the algorithm, MiniROCKET, was used. MiniROCKET produces approximately 10,000 features through random convolution kernels with varied parameters. It is fast compared to deep learning techniques and was shown to achieve state-of-the-art results over the public UCR timeseries dataset. While effective, this approach yields a high-dimensional feature set that introduces redundancy and increased storage demands. Moreover, this high-dimensional representation limits the available classifiers that can handle this vast number of features, limiting them to simple ones like the Ridge regression.
To address these issues, we developed a feature selection pipeline that includes feature representation using the Jeffries-Matusita distance and K-means clustering to select representative, non-overlapping features. The Jeffries-Matusita codes the pairwise class separation abilities of each feature. Application of K-means clustering to this new feature space groups similar features together, allowing to execute a simple and efficient sampling technique. Together, the proposed technique aims to maximize pairwise class separation while minimizing feature redundancy, in order to select a very small subset of features, as small as 1%.
A secondary goal of this work was to develop a measure that will support early classification in HAR, a procedure to determine the minimal signal length required for effective classification. Maximum Mean Discrepancy (MMD) was utilized to compare between the feature space of the full-length signals and their reduced versions. The feature-space representation was based on the MiniROCKET transform followed by the Jeffries-Matusita distance.
Experimental results were evoked on several HAR public datasets from the UCR repository, with an emphasis on drastic feature reduction. The proposed methodology was compared with other filter and wrapper-based feature selection techniques and showed competitive results. Moreover, we demonstrate the early detection technique and discuss the tradeoff between fast classification, lean data storage and classification accuracy.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום בצ'ט של שם מלא + מספר ת.ז.
