Personalising and predicting breakpoints for notifications on mobile devices

Abstract

Push notifications on smartphones can be disruptive due to inappropriate timing, leading to information overload and diminished user engagement. Turning off notifications increases stress for those with Fear of Missing Out (FoMO) and need to belong (NtB). Tailored solutions balancing connectivity and minimizing distractions are necessary. Determining optimal notification timing remains challenging due to the complexity of human behavior and lack of real-world data.

This thesis investigates Machine Learning (ML) and Deep Learning (DL) methodologies for predicting ideal mobile notification timings using physiological sensor data, psychological context, and situational context. The goal is to discern moments conducive to notification receptivity, elevating engagement while minimizing disturbances.

Analysis of the Ulster dataset shows statistical significance (p < 0.05) between user psychology, physiological data, and notification response times. An LSTM model achieves 98% R2 score, contrasting with 89% for ML techniques on the COMSATS dataset. Incorporating contextual parameters, the LSTM model achieves 21 minutes MSE in predicting breakpoint moments. However, a one-size fits-all model trained on AMI19 has over 60 minutes MSE, indicating the need for user-centric models.

A transfer learning strategy is introduced, transferring weights from a general LSTM model to personalized LSTM models, reducing training from 400 to 20 epochs. Resulting MSE scores range from 30-60 minutes, showing improved accuracy. But data scarcity and uneven distribution underlying this challenge. A CTGAN framework generates synthetic individual user data, addressing imbalances and refining MSE for tailored models.

Date of Award	Jun 2024
Original language	English
Sponsors	Invest NI - Competence Centre Program Grant RD0513853 - Connected Health Innovation Centre.
Supervisor	Ian Cleland (Supervisor), Christopher Nugent (Supervisor) & Paul Mc Cullagh (Supervisor)