CONTINUAL PERSONALIZATION OF SMARTPHONE-BASED ACTIVITY RECOGNITION ON DIVERSITYONE

Introduction

Smartphone sensors can be used for human activity recognition. DiversityOne is a large multi-country smartphone sensing dataset containing data from 26 sensor modalities and self-reports, making it a suitable benchmark for studying activity and behavior recognition in realistic everyday settings. This project investigates how activity-recognition models can be continually personalized to new users or contexts over time while reducing forgetting of previously learned knowledge.

Objectives

·      Study continual personalization for smartphone-based activity recognition on the DiversityOne dataset.

·      Compare static, fine-tuned, and continual-learning approaches.

·      Evaluate personalization gains, generalization, and forgetting across users or contexts.

Tasks

1.     Literature Review: Review smartphone-based activity recognition, personalization, and continual learning.

2.     Dataset Preparation: Use the DiversityOne dataset and define user- or context-incremental evaluation splits. DiversityOne includes smartphone sensing data collected with the iLog app, including accelerometer and gyroscope streams.

3.     Modeling: Train a baseline model for smartphone-based activity recognition from sensor data.

4.     Continual Personalization: Compare naive fine-tuning with two or more continual-learning methods.

5.     Evaluation: Assess accuracy, forgetting, and generalization across users, countries, or settings.

Pre-requisites

Python, machine learning, and interest in sensor data, mobile sensing, or continual learning.

Work

25% Theory, 55% Programming/Experiments, 20% Writing

Contact

Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)