Reproducing the Measurement of Filter Bubbles - Machine Learning Internship - Project 28

News recommendation systems (NRS) are becoming increasingly popular in news media to provide the most relevant news articles to the reader. However, there is a concern that personalized algorithms in news media could create filter bubbles, leading to less diversity in the material presented to users. We found a paper in which the existence of filter bubbles was analyzed on two aggregated and anonymized datasets provided by two Western European websites. In the study, it was proved that only a small decrease in topic variety in the first weeks after the sign-up process was present. Unfortunately, the original datasets used in the study were destroyed. After searching for new datasets, it became clear that NRS datasets of (personalized) recommendations with a satisfactory timespan and real users are not publicly available. After an intensive search, a dataset from the Danish news website "Ekstra Bladet" was chosen for further research. To deepen the research on filter bubbles, the given algorithm was also performed on a dataset of the Chinese short-video app "Kuaishou". We came to almost the same conclusion as the author but the results from both datasets were to our dissatisfaction inconclusive.