Learning in mobile context-aware applications
File(s)
Author(s)
Smith, Jeremiah
Type
Thesis
Abstract
This thesis explores and proposes solutions to the challenges in deploying context-aware systems that make decisions or take actions based on the predictions of a machine learner over long periods of time.
In particular, this work focuses on mobile context-aware applications which are intrinsically personal, requiring a specific solution for each individual that takes into account user preferences and changes in user behaviour as time passes.
While there is an abundance of research on mobile context-aware applications which employ machine learning, most does not address the three core challenges required to be deployable over indefinite periods of time. Namely, (1) user-friendly and longitudinal collection and labelling of data, (2) measuring a user’s experienced performance and (3) adaptation to changes in a user’s behaviour, also known as concept drift.
This thesis addresses these challenges by introducing (1) an infer-and-confirm data collection strategy which passively collects data and infers data labels using the user’s natural response to target events, (2) a weighted accuracy measure Aw as the objective function for underlying machine learners in mobile context-aware applications and (3) two training instance selection algorithms, Training Grid and Training Clusters which only forget data points in areas of the data space where newer evidence is available, moving away from the traditional time window based techniques. We also propose a new way of measuring concept drift indicating which type of concept drift adaption strategy is likely to be beneficial for any given dataset.
This thesis also shows the extent to which the requirements posed by the use of machine learning in deployable mobile context-aware applications influences its overall design by evaluating a mobile context-aware application prototype called RingLearn, which was developed to mitigate disruptive incoming calls. Finally, we benchmark our training instance selection algorithms over 8 data corpuses including the RingLearn corpus collected over 16 weeks and the Device Analyzer corpus which logs several years of smartphone usage for a large set of users. Results show that our algorithms perform at least as well as state-of-the-art solutions and many times significantly better with performance delta ranging from -0.2% to +11.3% compared to the best existing solutions over our experiments.
In particular, this work focuses on mobile context-aware applications which are intrinsically personal, requiring a specific solution for each individual that takes into account user preferences and changes in user behaviour as time passes.
While there is an abundance of research on mobile context-aware applications which employ machine learning, most does not address the three core challenges required to be deployable over indefinite periods of time. Namely, (1) user-friendly and longitudinal collection and labelling of data, (2) measuring a user’s experienced performance and (3) adaptation to changes in a user’s behaviour, also known as concept drift.
This thesis addresses these challenges by introducing (1) an infer-and-confirm data collection strategy which passively collects data and infers data labels using the user’s natural response to target events, (2) a weighted accuracy measure Aw as the objective function for underlying machine learners in mobile context-aware applications and (3) two training instance selection algorithms, Training Grid and Training Clusters which only forget data points in areas of the data space where newer evidence is available, moving away from the traditional time window based techniques. We also propose a new way of measuring concept drift indicating which type of concept drift adaption strategy is likely to be beneficial for any given dataset.
This thesis also shows the extent to which the requirements posed by the use of machine learning in deployable mobile context-aware applications influences its overall design by evaluating a mobile context-aware application prototype called RingLearn, which was developed to mitigate disruptive incoming calls. Finally, we benchmark our training instance selection algorithms over 8 data corpuses including the RingLearn corpus collected over 16 weeks and the Device Analyzer corpus which logs several years of smartphone usage for a large set of users. Results show that our algorithms perform at least as well as state-of-the-art solutions and many times significantly better with performance delta ranging from -0.2% to +11.3% compared to the best existing solutions over our experiments.
Version
Open Access
Date Issued
2015-08
Date Awarded
2016-02
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Dulay, Naranker
Sponsor
European Commission
Grant Number
264738
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)