Investigation of the consumption patterns on a media website

In order to implement an online recommendation system, the RTBF desires to analyze the users’ behavior on its website. This master thesis aims to achieve this purpose regarding the articles consumption on the RTBF website. A representation to visualize the one-step transitions within the website categories is firstly proposed in this work. Then, on the basis of several statistical tools (Maximum Likelihood Estimate, Akaike and Bayesian Information Criterion, Bayesian Method) and a cross validation method, this work exhibits that the knowledge of the current and previous article category for a randomly picked user is sufficient to predict his/her next one. Further, a clustering is performed for the users and a measure based on the Shannon entropy to compute the diversity of content read by the users is introduced. Eventually, an article recommendation algorithm is implemented and its performances are assessed. A global confinement phenomenon is observed in the users’ readings, people having a strong proclivity to restrict their consumption to a few categories of the website. It is also illustrated in this work that typical recommendation algorithms may reinforce this imprisonment phenomenon.