Probabilistic modelling and inference of human behaviour from mobile phone time series
Author(s)
Choujaa, Driss
Type
Thesis
Abstract
With an estimated 4.1 billion subscribers around the world, the mobile phone offers a unique
opportunity to sense and understand human behaviour from location, co-presence and communication
data. While the benefit of modelling this unprecedented amount of data is widely
recognised, a number of challenges impede the development of accurate behaviour models. In
this thesis, we identify and address two modelling problems and show that their consideration
improves the accuracy of behaviour inference.
We first examine the modelling of long-range dependencies in human behaviour. Human behaviour
models only take into account short-range dependencies in mobile phone time series.
Using information theory, we quantify long-range dependencies in mobile phone time series for
the first time, demonstrate that they exhibit periodic oscillations and introduce novel tools to
analyse them. We further show that considering what the user did 24 hours earlier improves
accuracy when predicting user behaviour five hours or longer in advance.
The second problem that we address is the modelling of temporal variations in human behaviour.
The time spent by a user on an activity varies from one day to the next. In order to
recognise behaviour patterns despite temporal variations, we establish a methodological connection
between human behaviour modelling and biological sequence alignment. This connection
allows us to compare, cluster and model behaviour sequences and introduce novel features for
behaviour recognition which improve its accuracy.
The experiments presented in this thesis have been conducted on the largest publicly available
mobile phone dataset labelled in an unsupervised fashion and are entirely repeatable. Furthermore,
our techniques only require cellular data which can easily be recorded by today's mobile
phones and could benefit a wide range of applications including life logging, health monitoring,
customer profiling and large-scale surveillance.
opportunity to sense and understand human behaviour from location, co-presence and communication
data. While the benefit of modelling this unprecedented amount of data is widely
recognised, a number of challenges impede the development of accurate behaviour models. In
this thesis, we identify and address two modelling problems and show that their consideration
improves the accuracy of behaviour inference.
We first examine the modelling of long-range dependencies in human behaviour. Human behaviour
models only take into account short-range dependencies in mobile phone time series.
Using information theory, we quantify long-range dependencies in mobile phone time series for
the first time, demonstrate that they exhibit periodic oscillations and introduce novel tools to
analyse them. We further show that considering what the user did 24 hours earlier improves
accuracy when predicting user behaviour five hours or longer in advance.
The second problem that we address is the modelling of temporal variations in human behaviour.
The time spent by a user on an activity varies from one day to the next. In order to
recognise behaviour patterns despite temporal variations, we establish a methodological connection
between human behaviour modelling and biological sequence alignment. This connection
allows us to compare, cluster and model behaviour sequences and introduce novel features for
behaviour recognition which improve its accuracy.
The experiments presented in this thesis have been conducted on the largest publicly available
mobile phone dataset labelled in an unsupervised fashion and are entirely repeatable. Furthermore,
our techniques only require cellular data which can easily be recorded by today's mobile
phones and could benefit a wide range of applications including life logging, health monitoring,
customer profiling and large-scale surveillance.
Date Issued
2009-09
Date Awarded
2010-03
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Dulay, Naranker
Sponsor
EPSRC
Creator
Choujaa, Driss
Grant Number
EP/C547705/1
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)