Supporting personalised recommendations in context-aware applications
Citation:
Shiu Lun Tsang, 'Supporting personalised recommendations in context-aware applications', [thesis], Trinity College (Dublin, Ireland). School of Computer Science & Statistics, 2009, pp. 219Download Item:
Tsang, Shiu Lun_TCD-SCSS-PHD-2009-06.pdf (PDF) 8.354Mb
Abstract:
Personalisation is the process of tailoring application behaviour to the requirements of its users. A primary concern of personalisation is providing users with behaviour recommendations that will aid them with their tasks. Making accurate recommendations benefits users by facilitating them with their activities, such as proposing actions preferred by the user, adapting resources to the user's situation, or informing users about news or events of interest. Accurate personalised recommendations also enable applications to make more effective use of user attention (e.g., by not interrupting users with irrelevant information) and facilitate greater user acceptance in proposed actions. In general, context-aware systems use information about the surrounding environment, known as context, to make behaviour decisions that aid users with their activities. Personalisation in existing context-aware systems works by adapting to a description of user preferences, which is either explicitly defined by domain experts or users or implicitly learnt by observing user behaviour. Both approaches are suited to small applications with well-defined domains, but are less suited to personalising larger context-aware applications for two main reasons. Firstly, both approaches rely on human input to identify, relate, and prioritise different contexts and user preferences (usually as rules or cost/similarity functions) to ensure correct behaviour. However, context-aware systems are likely to support a large set of contexts, implying a large set of context relationships and user preferences. The human task of accurately defining this set of information is correspondingly time-consuming, complex, and error prone. Secondly, existing approaches rely on the definition of algorithms that are inherently static.
Specifically, relevant information and the relative importance (or utility) of information, from which the preferred actions of users is inferred, are defined at development time. However, user preferences are likely to evolve with new experiences, and the mobility of context-aware application users means they are likely to encounter events, which represent new relationships between existing context that may affect their preferences. Changing user preferences and unconsidered context relationships will render pre-defined information inaccurate, and the static nature of current techniques means that applications are unable to automatically adapt to the changes in information that are required. Explicit user feedback is a possible solution to these challenges. However, explicit feedback is usually unsuitable due to the large cognitive load associated with ranking recommendation criteria or alternative behaviours. In addition, user mobility means that providing input when necessary is not always possible. Given the limitations of existing techniques, a new approach to personalisation for context-aware applications is therefore necessary - one which does not rely on static, development-time definition of information or require domain experts or users to explicitly specify and maintain preferences over time. This motivates the research question addressed by this thesis, which is: what techniques/algorithms are necessary to support the dynamic and implicit determination of user preferences from user behaviour, including relevant information and correct information utility, to facilitate context-aware applications in making accurate personalised recommendations to users?
To address this question, this thesis describes an investigation into a novel approach to personalisation that supports context-aware applications in making recommendations without explicitly predefining the set of context relationships, user preferences and recommended behaviour. Instead, the solution dynamically relates this information at run-time. It addresses the limitations of existing work by providing software components that facilitate user autonomy, unconsidered context relationships, and changing user preferences. At the core of the approach is the identification and examination of associations between contexts and user choices. These associations, generated from past user interactions, represent patterns in behaviour from which the preferred actions of users, for a particular context, are inferred. The main contribution of this thesis is the provision of a multi-staged recommendation process consisting of a set of operations and algorithms that: automatically elicits user preferences from user behaviour; dynamically filters relevant information and adjusts the relative utility of information to reflect the current context and up-to-date preferences of the user; and dynamically generates and ranks competing recommendations. The process uses existing techniques and algorithms (employed in various other fields), which are adapted and combined in a novel manner to provide an integrated approach for supporting personalisation in the context-aware domain.
The effectiveness of the approach in providing accurate personalised recommendations is evaluated by comparing recommended behaviour against the preferred choice of users under different contexts.
Two user studies and a set of computer simulated empirical experiments were conducted to assess accuracy. The results of two implemented recommendation strategies along with the accuracy measures for information filtering and utility adjustment algorithms are described. In addition, several of the most relevant recommendation approaches (rules, preference models, case-based reasoning, and neural networks) were implemented and evaluated, and the results of a study comparing the accuracy of these approaches against the solution described in this thesis are presented.
Sponsor
Grant Number
Irish Research Council for Science, Engineering, and Technology
Author: Tsang, Shiu Lun
Advisor:
Clarke, SiobhánQualification name:
Doctor of Philosophy (Ph.D.)Publisher:
Trinity College (Dublin, Ireland). School of Computer Science & StatisticsNote:
TARA (Trinity’s Access to Research Archive) has a robust takedown policy. Please contact us if you have any concerns: rssadmin@tcd.ieType of material:
thesisAvailability:
Full text availableLicences: