News recommendation with CF-IDF+

Abstract

Traditionally, content-based recommendation is performed using term occurrences, which are leveraged in the TF-IDF method. This method is the defacto standard in text mining and information retrieval. Valuable additional information from domain ontologies, however, is not employed by default. The TF-IDF-based CF-IDF method successfully utilizes the semantics of a domain ontology for news recommendation by detecting ontological concepts instead of terms. However, like other semantics-based methods, CF-IDF fails to consider the different concept relationship types. In this paper, we extend CF-IDF to additionally take into account concept relationship types. Evaluation is performed using Ceryx, an extension to the Hermes news personalization framework. Using a custom news data set, our CF-IDF+ news recommender outperforms the CF-IDF and TF-IDF recommenders in terms of F1 and Kappa.