Physical Activity Classification and Step Prediction

This repository contains part of the implementation developed during my thesis work. The project focuses on implementing and experimenting with several machine learning and deep learning models for physical activity classification and daily step prediction.

---

Project Overview

The project explores two related tasks using physical activity time-series data.

1. Physical Activity Classification

Different time-series classification models are used to classify activity patterns.

Implemented models include:

• ROCKET
• MiniRocket
• MiniRocketVoting
• Multi-Layer Perceptron (MLP)

2. Daily Step Prediction

Time-series forecasting models are used to predict future daily step counts.

Implemented models include:

• Bidirectional LSTM
• GRU
• 1D Convolutional Neural Network (CNN-1D)

---

Dataset

The experiments use physical activity datasets derived from the PMData and the MOX2-5 dataset.

Main features used in the experiments include:

• steps — daily step count
• sedentary — sedentary time (minutes)
• LPA — light physical activity (minutes)
• MPA — moderate physical activity (minutes)
• VPA — vigorous physical activity (minutes)

Additional sleep-related features such as sleep duration and sleep score are also used.

Processed datasets used in the experiments are included in this repository.

---

Repository Structure

0/
Data preprocessing for the PMData dataset. Raw Fitbit activity and sleep data are aggregated at the daily level and saved as a processed dataset.

1/
Label generation for the classification task. Activity labels are created based on step counts and activity intensity features.

2/
Exploratory data analysis and feature analysis including correlation analysis, class distribution visualization, feature selection, PCA, and feature importance analysis.

3/
Implementation and evaluation of classification models including ROCKET, MiniRocket, MiniRocketVoting, and MLP. Results include confusion matrices and training performance plots.

4/
Time-series forecasting experiments for daily step prediction using deep learning models such as LSTM, GRU, Bidirectional LSTM/GRU, stacked architectures, and a 1D CNN.

plotting/
Visualization and exploratory analysis of the datasets including feature distributions, activity level distributions, and participant-level analysis.

---

Results

Classification Results

The classification experiments were evaluated using confusion matrices for the following models:

• ROCKET
• MiniRocket
• MiniRocketVoting
• Multi-Layer Perceptron (MLP)

Generated confusion matrices and training plots for both the PMData and MOX datasets are included in the repository.

---

Step Prediction Results

The following table summarizes the forecasting performance of different deep learning models for daily step prediction.

PMData Dataset

Model	RMSE	MAE	RSD	ET
CNN-1D	1164	230	11656	1288
Bidirectional GRU	4468	3359	4470	1372
Vanilla LSTM	4633	3419	4636	1412
Stacked LSTM	4662	3461	4664	3219
Stacked GRU	4788	3546	4791	3706
Vanilla GRU	4815	3583	4818	1253
Bidirectional LSTM	5006	3776	5009	2252

MOX Dataset

Model	RMSE	MAE	RSD	ET
CNN-1D	1510	413	1513	108
Stacked GRU	4767	3534	4778	376
Stacked LSTM	4798	3569	4809	414
Vanilla GRU	4836	3578	4847	179
Bidirectional LSTM	4851	3585	4862	349
Bidirectional GRU	4865	3589	4867	310
Vanilla LSTM	4881	3619	4893	190

Lower values of RMSE and MAE indicate better prediction performance.

Requirements

Python 3 is required.

Install dependencies with:

pip install -r requirements.txt

Main libraries used include:

• pandas
• numpy
• scikit-learn
• tensorflow / keras
• sktime

---

Reference

This project was developed as part of my thesis work and includes implementation and experimentation based on methods described in the following research paper:

An automatic and personalized recommendation modelling in activity eCoaching with deep learning and ontology

The repository focuses on implementing parts of the data processing and modeling pipeline and experimenting with several machine learning models for activity classification and step prediction.