Abstract
Context-aware information retrieval helps providing resources that are relevant to the current situation of users. In many work environments, carrying out a task demands massive information access and management. Context-aware computing in such settings can reduce the need for manual information retrieval and assist the users in bringing forward relevant resources and hence decreasing the time and effort needed to manage and update the information. So far, research in context-aware computing has mainly focused on adapting the devices and systems to the individual user’s need or behaviour. The system detects different patterns of a user’s work and adapts accordingly. However, in a collaborative work environment such as a hospital, adaptive information retrieval goes beyond covering the need of individuals, as clinical tasks often involve several users working concurrently sharing systems, devices and documents. This thesis proposes an approach to the problem of context-aware information retrieval in multi-user collaborative environments with public shared devices. We identify three sub problems to be addressed; 1) detection of the situation in which the device is present; 2) recognition of sequential multiple concurrent activities in the observable space of the device; and 3) adaptation of the device to present the relevant information to current situation. For the problem of detecting the situation of a device, we argue that in order to provide more precise information adaptation on different devices spread in a ubiquitous shared environment, the context of a device should be weighted. This can be done by recognizing not only the contextual elements related to that device (e.g., location and time) but also the situation pattern in which the device is present. For this, we introduce a method called ‘L-P-A Walk’ that identifies such situation patterns in the observable space of a device and helps weighting context elements that matter for that situation. For the problem of activity recognition, we address the issues of i) multiple concurrent activities and ii) sequential dependency between activities. For the former, we propose joint- and parallel-learning mod els; and to address the latter, we add virtual and historical evidence as features to the learning models. We perform classification on the data and evaluate the output by proposing a distance based method. The best performance is obtained in parallel learning with historical evidence. For the problem of information adaptation, we point to the fact that most real world situations have been experienced at least once before either in the same or a different setting. We argue that despite slight differences, there is always a most similar situation to the current one that can provide an initial set of handling or actions that might also be relevant to the new situation. This initial set is then updated, evolved, and adapted to the specific characteristics/profile of the current situation. By incorporating collaborative filtering, we dynamically provide most relevant types of information that has been used in similar situations. An evaluation of our approach with clinicians shows that the proposed mechanism is able to i) incrementally build and update the model of relevant information for every situation based on similar past cases and ii) incorporate users’ information choices as implicit feedback and retrain the model to provide more satisfactory information assistance.
Originalsprog | Engelsk |
---|---|
Kvalifikation | Doktor i filosofi (ph.d.) |
Vejleder(e) |
|
Status | Udgivet - 2011 |