Content and Type as Orthogonal Modeling Features: a Study on User Interest Awareness in Entity Subscription Services

Real-word entities can be mapped to unique entity identifiers through an Entity Name System (ENS), to systematically support the re-use of these identifiers and disambiguate references to real world entities in the Web. An entity subscription service informs subscribed users of changes in the descriptive data of an entity, which is a set of free-form attribute name-value pairs. We study the design, implementation and application of an adaptable, push-policy subscription service, within a large-scale ENS. The subscription system aims to deliver ranked descriptions of the changes on entities, following user preferences through a feedbackdriven adaptation process. Within this paper, we offer a novel approach based on the discrimination between the content (descriptive aspect) and the type (directly quantifiable or binary aspects) of information instances (i.e., entity changes in our case). We study and evaluate two different approaches of adaptation to user interests: one that only manages constant user preferences and one that adapts to interest shifts of the users. We evaluate the learning curve of the variations of the system, the utility of the content-type discrimination, as well as the effectiveness of modeling user behavior for random interest shifts of the user. The experiments demonstrate good results, especially in the system’s content-aware adaptation aspect, which takes into account user interest shifts. We also extract a set of useful conclusions concerning the detection of user interest shifts, based on feedback, as well as the relation between the user interest shift frequency and the optimality of learning memory window size.