Recommender Systems For The Semantic Web

Recommender Systems for the Semantic Web
Antonis Loizou 1 and Srinandan Dasmahapatra 2
Abstract. This paper presents a semantics-based approach to Recommender Systems (RS), to exploit available contextual information about both the items to be recommended and the recommendation process, in an attempt to overcome some of the shortcomings of traditional RS implementations. An ontology is used asa backbone to the system, while multiple web services are orchestrated to compose a suitable recommendation model, matching the current recommendation context at run-time. To achieve such dynamic behaviour the proposed system tackles the recommendation problem by applying existing RS techniques on three different levels: the selection of appropriate sets of features, recommendation model andrecommendable items. changeable to the ones rated highly by users, ignoring potential user requirements. The two approaches are often combined in Hybrid RS to achieve improvements in the quality of recommendations [1, 3]. These shortcomings reflect the lack of computational support for humans who are interested in items they, or the people who usually share their taste haven’t previously come across. Inaddition, such systems do not allow for shifts of the user’s interest over time, since all ratings provided by a user have an equal bearing on the recommendation selection. To clarify this point consider the following conceptualisation: A user X has provided high ratings only for items in some set A, however (s)he is now only interested in items from another set, B. A conventional RS will not beable to recommend items from set B until enough ratings are provided for items in B, in order for them to dominate in the clustering and selection processes. This means that a system shouldn’t become stable, and that the classification of the same items to different classes, at different times, may be deemed correct, something that would be unacceptable in most machine learning contexts. To accountfor this requirement of time dependance on users’ preference context, conventional architectures recompute their user clusters periodically, effectively choosing a different training set every time. This can aggravate problems caused by data sparsity, and important modelling decisions about transitions between user needs have to be addressed. Furthermore, while it is apparent that an artifact’sfeatures have a bearing on whether it appears interesting or not, users may not be able to identify its desirable characteristics at the outset. For instance, someone who wants to buy a new car might only specify ”I want a black car” to begin with. Instead of buying the first black car available, s/he might look at a variety of black cars and as their knowledge of cars grows in the process, discoverother possible features of interest, or even come across an unusual opportunity and end up buying a different coloured car. This would suggest that for a RS to be successful, it needs to be able to identify which of an item’s features may potentially be of interest to the user, against a variety of possible modes of generalisation. To overcome such issues, a system should be able to consider thesemantics of both the recommendation context and those of the items at hand to constrain the recommendation process. Information specific to the recommendation context for both user clustering and content-based comparisons have been shown to improve overall recommendation performance [3, 17, 21]. By incorporating relevant contextual information into a recommendation model, we enable the system toevaluate the appropriateness of a given recommendation based on some heuristics, for example the time of recommendation or the utility of the recommended item to the user [1, 6, 9]. The system proposed in this paper is one designed to choose appropriate input and output spaces dynamically, in a manner that will allow for real time recommending, matching the variable temporal and contextual...