Time-Series Compression for Classification: Performance, Efficiency and Robustness on FordA and ECG5000

Time-Series Compression for Classification: Performance, Efficiency and Robustness on FordA and ECG5000

Computer Science – Machine Learning and Time-Series Analysis

Kaitao (Kenneth) Jiang

Weiren Yu

This project investigates whether time-series signals can be compressed without losing the information needed for reliable classification. Four compression methods—PAA, PLA, APCA and Fourier representations—are evaluated on the FordA and ECG5000 datasets using a shared neural-network classifier. The experiments examine classification performance, training and inference efficiency, and robustness to noise and missing values when only 50% or 25% of the original input width is retained. Signal reconstructions and PCA visualisations are also used to examine which structures remain after compression. The results show that Fourier representations provide the strongest overall trade-off, sometimes improving classification performance while reducing input width and model size, although computational speed-ups depend on the dataset and compression method.

machine learning, dimensionality reduction, signal processing, Fourier transform, neural networks, robustness