Semantics-based news recommendation with SF-IDF

Content-based news recommendations are usually made by employing the cosine similarity and the TF-IDF weighting scheme for terms occurring in news messages and user profiles. Recent developments, such as SF-IDF, have elevated news recommendation to a new level of abstraction by additionally taking into account term meaning through the exploitation of synsets from semantic lexicons and the cosine similarity. Other state-of-the-art semantic recommenders, like SS, make use of semantic lexicon-driven similarities. A shortcoming of current semantic recommenders is that they do not take into account the various semantic relationships between synsets, providing only for a limited understanding of news semantics. Therefore, we extend the SF-IDF weighting technique by additionally considering the synset semantic relationships from a semantic lexicon. The proposed recommendation method, SF-IDF+, as well as SFIDF and several semantic similarity lexicon-driven methods have been implemented in Ceryx, an extension to the Hermes news personalization service. An evaluation on a data set containing financial news messages shows that overall (by accounting for all considered cut-off values) SF-IDF+ outperforms TF-IDF, SS, and SF-IDF in terms of F1-scores.