Documents, Topics, and Authors: Text Mining of Online News

Abstract

The goal of recommender systems is, in essence, to help people to discover items they might like, i.e., items that fit their preferences, personality, and needs. Depending on the respective domain, those items can be books, movies, music, hotels, and much more. Typically, recommendations are based on past user interactions (e.g., movies a user saw, hotels a user booked, etc.). This work in progress paper focuses on news recommender systems. Because of the nature of news (e.g., constantly new items, short item lifetime, etc.), recommendations based on past interactions are especially hard to make. Hence, news recommender systems heavily rely on the actual content of news. While previous work mainly considers one aspect of the content of news articles, we jointly analyse and discuss in this work a given corpora of news articles on three different levels (i.e., document-level, topic-level, and author-level). The overall aim is to set to provide the basis for a comprehensive news recommender system, which reaches beyond accuracy and considers also diversity and serendipity. We demonstrate that relevant information can be extracted out of a given corpora, and differences in author, time, and topic can be shown. Furthermore, the author-level analysis shows that documents can be clustered based on the writing style of authors. Finally, our findings show that author-level analysis has the potential to recommend the most diverse items compared to the other approaches.