JiT-PIe: Just-in-Time Personalisation of Internet and e-content

The rapid growth of the Internet, wireless communication, multimedia home and office servers, and the data interoperability among massive digital repositories will enable virtually everyone to have access to huge amounts of information on almost any topic. Up to now, the vast majority of the technology has been directed towards collecting information and providing access to various information resources. It is now obvious that we must redirect at least some part of our research to providing technologies for developing relevant and, to the extent possible, complete personal information spaces; visualisation methods that enable users to process in parallel vast quantities of information; and interaction paradigms that allow an interaction that will be closer to cognitive processes. This project proposes an attempt towards the development of tools and techniques for constructing just-in-time a personal information space in the form of a personal Internet.

The development of this personal Internet in which both e-content and navigation, as well as the user interface are adapted differs in many respects from applications of computational intelligence and machine learning methods like data mining, and knowledge extraction and hence leads to many new challenges. The first challenge is to recast the problem of user modelling in terms of a learning task. Nowadays, static user modelling as used for information filtering and recommendation usually attempts to place content in two categories: interesting, uninteresting. Clearly this approach fails to capture the variety in user interests, the time-varying nature of the user characteristics and the shift of interest from one type of content to another one. The second challenge has to do with the encoding of the user data and user models in a way that induction is tractable, decisions are clearly interpretable and personalisation of content and navigation is supported. A third challenge is concerned with gathering data for training. In this context, interaction with the user is the primary source of information. For example, explicit or implicit user feedback can be used for learning to recommend information according to user interests. An important difference between learning methods suitable to work in this context and other, general learning algorithms lies in our need for on-line learning. Although, in some cases, this is not a strict requirement, in that data can be collected during the interaction with the user and the learning algorithm can be run off-line, it is acknowledged that the on-line approach is more desirable. Furthermore, training data are quite precious, in the sense that as opposed to other data mining and knowledge extraction applications, we want to generate an accurate model of the user from a rather limited number of training cases, due to restrictions placed on the time users can spend interacting with the system. This implies the need for learning methods, which are suitable for on-line training, can achieve high generalisation from small training sets, and satisfy the requirement for fast and accurate learning which this projects aims to achieve.

Project Team: G.D. Magoulas and D. Dimakopoulos (Birkbeck); M.C. Angelides (Brunel)

Support: E.P.S.R.C. (GR/R92554/01-02)

Duration: 2002-2005

Home  - Teaching  -  Research  -  Publications  -  Bio  -  Links