Prepare temporal sequences
Work with app usage records from approximately 12,000 users, including discretisation, numericalisation and padding.
Master's thesis · 2022
Forecasting how people use their smartphones.
Using large-scale behavioural data and a hierarchical attention-based deep learning approach to model app usage patterns and predict weekly activity.
Usage data is sequential and deeply personal.
People interact with many apps over the course of a day. Those interactions change over time, producing complex behavioural sequences rather than simple, independent observations.
The research explored whether these patterns could be modelled to forecast weekly smartphone activity and reveal more about how usage evolves over time.
From raw behaviour to learned patterns.
The study brought together data preparation, sequence modelling and comparative evaluation. A hierarchical attention network (HAN) formed the core of the proposed approach.
Learn with attention
Explore hierarchical attention-based models with embeddings, positional information and sequence encoders.
Compare meaningful baselines
Evaluate the proposed models against persistence and simpler GRU/LSTM approaches, then examine attention weights.
What the research showed.
Better than the tested baselines
The original project reports that HAN-based models outperformed persistence and simple GRU/LSTM baselines in the evaluated classification tasks.
Patterns across time scales
The hierarchical approach captured information about both daily and weekly patterns in smartphone behaviour.
More than a prediction
Exploring attention weights offered an additional way to investigate what signals the learned models were responding to.
Research discipline that still translates.
This project reflects an approach that remains useful far beyond smartphone data: understand the structure of a problem, prepare trustworthy input, compare against credible baselines, and interpret results before declaring success.
It also established a foundation in sequence modelling and attention-based architectures, areas that connect naturally to more recent AI engineering work.
Want the technical detail?
The full research covers the experimental design, model architecture, evaluation and additional findings.