{mns2, ha} @ ecs.soton.ac.uk Abstract. Many advanced recommendation frameworks employ ontologies of various complexities to model individuals and items, providing a mechanism for the expression of user interests and the representation of item attributes. As a result, complex matching techniques can be applied to support individuals in the discovery of items according to explicit and implicit user preferences. Recently, the rapid adoption of Web2.0, and the proliferation of social networking sites, has resulted in more and more users providing an increasing amount of information about themselves that could be exploited for recommendation purposes. However, the unification of personal information with ontologies using the contemporary knowledge representation methods often associated with Web2.0 applications, such as community tagging, is a non-trivial task. In this paper, we propose a method for the unification of tags with ontologies by grounding tags to a shared representation in the form of Wordnet and Wikipedia. We incorporate individuals ’ tagging history into their ontological profiles by matching tags with ontology concepts. This approach is preliminary evaluated by extending an existing news recommendation system with user tagging histories harvested from popular social networking sites.

Abstract. Personalised recommender systems learn about a user’s needs, and identify and suggest information items (news articles, images, videos, etc.) that meet those needs. User needs can be explicitly or implicitly defined either in the form of user tastes, interests and goals, or by system parameters and configurations. Most research efforts in the Recommender Systems field can be said to have been directed towards either defining and improving techniques that provide item recommendations from available preference data, or defining techniques for learning the latter. However, little research has focussed on learning which preferences are really relevant to provide accurate recommendations, and which ones imply anomalous behaviour of the recommendation mechanisms. We present a meta-evaluation methodology that applies Machine Learning techniques to analyse log information of a personalised news recommender system in order to discover (and rank) which user preferences and system settings are suitable for accurate recommendations. We also show how the proposed methodology can be used to ease the system evaluation itself.

Abstract. Recommender systems suggest users information items they may be interested in. User profiles or usage data are compared with some reference characteristics, which may belong to the items (content-based approach), or to other users in the same context (collaborative filtering approach). These items are usually presented as a ranking, where the more relevant an item is predicted to be for a user, the higher it appears in the ranking. In this scenario, a preferential order has to be inferred, and therefore, preference learning methods can be naturally helpful. The relevant recommendation model features for the learningbased enhancements explored in this work comprise parameters of the recommendation algorithms, and user-related attributes. In the researched approach, machine learning techniques are used to discover which model features are relevant in providing accurate recommendations. The assessment of relevant model features, which is the focus of this paper, is envisioned as the first step in a learning cycle in which improved recommendation models are produced and executed after the discovery step, based on the findings that result from it.

The elements that can be considered under the notion of context in a recommender system are manifold: user tasks/goals, recently browsed/rated items, computing platforms and network conditions, social environment, physical environment and location, time, external events, etc. Complementarily to these elements, we propose a particular notion of context for semantic content retrieval: that of semantic runtime context, which we define as the background topics under which activities of a user occur within a given unit of time. A runtime context is represented in our approach as a set of weighted concepts from domain ontologies, obtained by collecting the concepts that have been involved in user’s actions (e.g., accessed items) during a session. Once the context is built, a contextual activation of user preferences is achieved by finding semantic paths linking preferences to context. In this paper, we present a user-centred study of our context-aware recommendation model using a news recommender system called News@hand. We analyse the strengths and weaknesses of our approach, and discuss the importance of contextualisation in a news recommendation scenario.