Time Series User Classification

machine-learning
time-series
python
classification
User identification from walking activity using accelerometer data — comparing Random Forest vs ROCKET classifiers
Published

July 14, 2021

Overview

Multiclass classification project identifying users from smartphone accelerometer data collected during walking activity (22 participants). Compares traditional feature engineering with time series windowing (Random Forest) against state-of-the-art ROCKET (Random Convolutional Kernels) classifier.

Accelerometer data view

Key results:

Classifier Test Accuracy Balanced Accuracy
RF Window 50 0.61 0.54
ROCKET Window 100 0.70 0.61

ROCKET significantly outperforms Random Forest using only raw x, y, z features — no manual feature engineering required.

Dataset: UCI — User Identification From Walking Activity

Code

GitHub Repository