Modelos estocásticos para leitores de jornais online

Abstract

The aim of this thesis is the study of the behavior of online users of digital newspapers. We analyzed more than 20 million sessions composed by users successive clicks in news posted in two large Brazilian online newspapers. The motivation for this work isthat understanding the sequence of topics reading behavior can help to design better recommendation systems. Each user session was reduced to the sequence of topics read. We analyzed 32 stochastic models, each one trying to capture the essence of the userbehavior. They are divided into five categories: models without past influence, those that totally disregard the information of the past; short memory models, where only the recent topics read affect the next one; preference revealed models which the future is conditioned on characteristics of a topic at a time; geometric permanence modelswhere the reading behavior is divided among the options of remaining on the current topic of reading following a geometric distribution and changing of topic according to some rules; and finally models of cumulative advantage, in which previous readings of a topic increase its readings chances in the future. The models are fitted by maximumlikelihood and compared according to goodness of fit and prediction power. The best models are those in which the user moves around the states influenced by his most recent readings. The cumulative advantage models were close behind, with slightly worse predictions but still quite satisfactory. We show how our findings can be explored for dynamically recommending online news to a user based on his clicks tracking in agiven reading session.