Metrical Representation of Readers and Articles in a Digital Newspaper

Abstract

Personalized recommendation of news in digital journals have to deal with important peculiarities. A majority of users (readers) are anonymous, and frequently news are volatile, they have an extremely short duration while other items arise. In this paper, we learn a mapping of users and items into a common Euclidean space where the similarities can be computed in a linear geometric context. The location of readers in the map are re ned as they read more articles, and at the same time news can be inserted or removed eas- ily. The metric properties of readers and news will pave the way for a solid base to o er recommendations for readers not only adjusted to their tastes, but with a certain degree of di- versity or serendipity. Additionally, clusters of readers with similar interests or tastes could be discovered and exploited for marketing purposes. This mapping is learned using a scalable factorization algorithm that aims at optimizing the accuracy of the personalized recommendations. The paper includes an experimental study done with real word data