Recommendation Strategy based On Relation Rule Mining
Mehdi Adda, Chabane Djeraba · 2005
Web users are nowadays confronted with the huge variety of available information sources whose content is not targeted at any specific group or layer. Recommendation systems aim at adapting this content to (their guesses about) the needs of a particular user and hence usually compute some sort of relevance score of the manipulated content objects. As direct information about user needs is scarce, content objects are assessed not directly with respect to those needs but rather in relative manner, i.e., as compared to other objects whose relevance is known. The likeness indices for objects vary from association degrees computed from user logs to inter-object similarities to aggregations of direct user votes on object relevance. We claim that as structured content descriptions, i.e., by means of an ontology, get ever more popular among information providers on the Web, the underlying domain knowledge may successfully be exploited in comparing objects for recommendation purposes. In this paper, we introduce a recommendation approach that explores a specific sort of domain knowledge, the inter-object relational links (e.g., part-of, powered-by, same-author-as, etc.), that are typically expressed at the ontological level by means of specialized languages like OWL. These links form the backbone of a new sort of behavioral patterns, called relation rules, that are extracted from user logs. The basic notions, definitions and mining algorithm for relation rules are provided and illustrated by means of sample ontology and content object set of e-commerce flavor. Key-Words: Recommendation systems, personalization, ontology, data mining. 1